(Bradford Books) Manfred Fahle, Tomaso Poggio - Perceptual Learning-MIT Press (2002)
(Bradford Books) Manfred Fahle, Tomaso Poggio - Perceptual Learning-MIT Press (2002)
(Bradford Books) Manfred Fahle, Tomaso Poggio - Perceptual Learning-MIT Press (2002)
A Bradford Book
The MIT Press
Cambridge, Massachusetts
London, England
( 2002 Massachusetts Institute of Technology
Introduction ix
II Low-Level Psychophysics 145
IV Modeling 335
Glossary 381
References 387
Contributors 443
Index 447
Introduction
Plasticity is defined here as modifiability of the Figures (Gollin 1960), and when perceiving the
brain leading to more appropriate function. The Dalmatian dog or depth in random-dot stereograms
term most often describes the neuronal substrate of ( Julesz 1971), can last for months and should rather
changing behavior, such as changes in synaptic be considered as perceptual learning.
weights or formation of new synapses. Plasticity, for Development and maturation are also coupled with
example, serves to adjust the functional and ana- changes of (cortical) information processing and of
tomical organization of the central nervous system behavior. These terms, unlike learning, ascribe the
as a result of sensory experience or other factors. main thrust of the changes in a behavior to genetics,
The term also describes functional and anatomical not the environment. Hence, changes in observable
changes by which the central nervous system alle- behavior are seen as the consequences of the grow-
viates limitations caused by lesions. The term adap- ing and maturation of the organism along a largely
tation is most often used in relation to changes of predefined path rather than as a consequence of
information processing on relatively peripheral levels information gathered through interaction with the
in the central nervous system, as a result of extended outside world (in practice, of course, genetics and
presentation of a stimulus, or after a change in environmental influences strongly interact in deter-
environmental conditions. A classic example is light mining behavior).
adaptation. There, the working range of the visual Improvement through insight, as the name indicates,
system is shifted to the most appropriate level to deal is a term that should be reserved for more e¤ective
with the changes in ambient light intensity. Hence, information processing based on cognitive processes
adaptation usually means an adjustment within a and incorporating an abrupt change, such as one-
predefined working range, with no long-term changes shot learning does (cf. also chapters 13, 14). Such
of the nervous system. (In psychology, ‘‘adaptation’’ processes play a role also in perceptual learning
tends to be used also for long-lasting changes.) Ha- but come close to declarative forms of learning.
bituation (called satiation in the context of rewards) Memory usually is a consequence of learning. It im-
seems to be a special case of adaptation, namely a plies the coding, storage, and retrieval of informa-
shift of working range towards lower sensitivity, as in tion. While perceptual learning certainly is based on
the case of decreased reflex response after repeated some form of information storage, this information is
stimulation while sensitization indicates a temporal in- stored implicitly by changing the way the stimuli are
crease in sensitivity, for example, after a painful stimu- analyzed.
lus. Aftere¤ects can be considered, in many cases, as
the result of adaptation, most often habituation, in
Stimulus Specificity of Perceptual Learning
cases where perception is the result of antagonistic
Causes a ‘‘Paradigm Change’’
neuronal channels. Duration of aftere¤ects is usually
short. Priming describes the e¤ect of a (sometimes sub-
Visual recognition of objects improves through
liminal) stimulus on subsequent perception of (other)
training, and ‘‘practice makes perfect’’ (Volkmann
stimuli and/or behavioral responses. The e¤ect of
1863; Best 1900). For beginners in the study of his-
priming is usually short, vanishing after a few tens of
tology, all specimens look quite similar (for example,
seconds, while some e¤ects usually called priming,
liver, lung, or kidney). Sooner or later, however, the
such as the ones obtained by means of the Gollin
xii Introduction
advanced student wonders how one could possibly obtained through training was specific for stimulus
miss the di¤erence. This type of visual classification orientation (cf., however, Ball and Sekuler 1987).
is a relatively complex visual task, but it has also been A similar specificity of improvement even for the
known for decades that performance in much sim- retinal position and the eye used during training was
pler perceptual tasks improves through practice, as observed in a texture discrimination task where
for example vernier discrimination (McKee and observers had to discriminate a figure from its sur-
Westheimer 1978), stereoscopic depth perception round based on the orientation of the stimulus ele-
(Ramachandran and Braddick 1973), and discrimi- ments (Karni and Sagi 1991; chapter 10). Moreover,
nation between line orientations (Vogels and Orban improvement did not transfer between a three-dot
1985). Generally, however, these improvements in vernier and a three-dot bisection task, although the
discriminative power were not attributed to changes stimuli of both tasks di¤er by approximately one
on the peripheral or early levels of information pro- photoreceptor diameter only, thus excluding all
cessing, but were, implicitly or even explicitly, con- explanations based on better accommodation or fix-
sidered to be based (exclusively?) on cognitive, that ation (Fahle and Morgan 1996). Hence, the changes
is, high-level changes of visual information processing. of the nervous system underlying this form of per-
During the last decade, however, a number of ceptual learning should occur on a quite early level
electrophysiological and psychophysical experiments of cortical information processing where the neurons
have shed some doubts on this purely cognitive in- are already selective for di¤erent stimulus orientations
terpretation, suggesting that even the adult primary (unlike in the retina), but still partly monocularly
sensory cortices show much more plasticity than was activated unlike in all cortical areas beyond the pri-
hitherto believed. The first evidence of this was the mary visual cortex. In particular, the eye specificity
specificity of improvement through learning for of the improvement as well as its high positional and
stimulus orientation. Fiorentini and Berardi (1980; orientation specificity (less than 5-degree bandwidth,
chapter 9, this volume) were the first to describe cf. chapter 11) point to the primary visual cortex as
such orientation specificity, finding that practice the most probable site for at least a large part of the
improved discrimination between complex gratings, changes underlying this form of perceptual learning.
but that the improvement did not transfer to stimuli Perceptual learning, however, is not restricted to
rotated by 90 degrees. Poggio, Edelman, and Fahle these rather simple visual stimuli, but extends to
(1991; chapters 11, 15, 18) described a similar ori- more complex line figures (chapter 15), simple ob-
entation specificity for a vernier discrimination task jects (chapter 16) and faces (chapter 17). Training ob-
where observers had to indicate whether the right servers with repeated presentations of iso-oriented
segment of a horizontal vernier was above or below ( parallel) lines increases detection sensitivity for these
the left segment, or whether the lower segment of a stimuli, probably by strengthening lateral inter-
vertically oriented vernier stimulus was o¤set to the actions that exist in the primary visual cortex be-
right or to the left relative to the upper segment. tween the corresponding neurons (von der Heydt
Both groups of observers improved performance and Peterhans 1989; Dresp and Bonnet 1991; Polat
highly significantly during training but their perfor- and Sagi 1994; Dresp 1999; chapter 10).
mance returned to base level when the stimulus was Recent neuroanatomical (see chapter 1) and elec-
rotated by 90 degrees. Hence, the improvement trophysiological evidence (chapters 2–7) supports
Introduction xiii
this conclusion. The first indications for plasticity in they were trained with (Kobatake, Wang and Tanaka
the wiring of the adult primary sensory cortex stem 1998; Tanaka, chapter 4, this volume; Logothetis,
from the somatosensory cortex, where Merzenich Pauls, and Poggio 1995; Sheinberg and Logothetis,
and collaborators found an amazing amount of chapter 6, this volume). The e¤ects of structural
change in the distribution of receptive fields after lesions of the retina can be mimicked by presenting
lesions (Merzenich et al. 1988; see also chapters 2 an artificial scotoma, that is, by preventing visual in-
and 7). And in the somatosensory system, perceptual put to part of the retina. The consequences are sim-
improvement was associated with a highly selective ilar to if less pronounced than those of structural
cortical reorganization in the primary somatosensory lesions (Volchan and Gilbert 1994). Presenting one
cortex, S1, with high stimulus selectivity (Recan- stimulus within a cortical receptive field and another
zone et al. 1992; and chapter 2). Single-cell record- one simultaneously close by produces expansion of
ings in adult monkeys demonstrated that receptive the receptive field towards the outside stimulus
fields in the primary visual cortex can change posi- (Eysel, Eyding, and Schweigart 1998; see chapters 4
tion after retinal lesions (Gilbert and Wiesel 1992; and 6). Similar, but far more pronounced changes
Eysel and Schweigart 1999; chapter 3). Apart from occur after coactivation in the somatosensory cor-
their location, receptive fields also can radically tex (Diamond, Armstrong-James, and Ebner 1993;
change their receptive field sizes within minutes after Wang, Merzenich, Sameshima, and Jenkins 1995;
retinal lesions. The increase in size is about an order Godde, Stau¤enberg, Spengler, and Dinse 2000),
of magnitude for cortical receptive fields close to the indicating that the extent of adult plasticity may dif-
lesion’s border (Gilbert and Wiesel 1992). Accord- fer between the two systems. As a consequence of
ingly, the distribution of mass potentials evoked by these experimental results, we now know for sure
visual stimulation in humans changes as a result of that even the adult primary sensory cortices have a
training, especially pronounced for short latencies fair amount of plasticity to perform changes of in-
over the primary visual cortex (Fahle and Skrandies formation processing as a result of training. These
1994). Even more complex single cell properties results have far-reaching implications for the concept
change as a function of training, and shifts of recep- of memory, too. ‘‘Memory’’ in perceptual learning is
tive field position, which amount to about 2 mm of not a property of one or a few assemblies of cortical
cortical distance for short-term learning, can increase neurons in specialized parts of the brain, but all cor-
to about 8 mm for long-term learning (Pons et al. tical areas seem to be able to change and adapt their
1991). Animals trained to discriminate pitches around function both on a fast and a slow time scale.
a defined standard frequency improve specifically
for this frequency band, and the corresponding
Multiple Levels of Learning and Top-Down
representation in primary auditory cortex increases
Influences
(Recanzone, Schreiner, and Merzenich 1993). Train-
ing to discriminate the main motion direction in
The earlier insight is still true that changing the very
dynamic random dot patterns improves both perfor-
front end of information processing as a result of
mance of the animal and the response characteristics
learning one perceptual task would necessarily
of single cells in area MT (Zohary et al. 1994), and
change processing of many other stimuli presented
neurons in inferotemporal cortex change their re-
to the same sensors. The speed and amount of learn-
ceptive field properties according to complex stimuli
xiv Introduction
ing depends strongly on attentional control (Ahissar eye trained for exact visual field position and for
and Hochstein 1993; chapter 14), as well as on ‘‘in- stimulus orientation (Ahissar and Hochstein 1997;
sight’’ (Rubin, Nakayama, and Shapley 1997; chap- chapter 14).
ter 13)—that is, on top-down influences within the To conclude, perceptual learning di¤ers from
brain. Receptive fields of cortical neurons even in other forms of learning in that it can be very task-
the primary visual cortex change in a task-dependent and stimulus-specific and probably involves func-
way, influenced by the actual perceptual task pre- tional and anatomical changes even in primary
sented to the animal (Gilbert 1998; Gilbert et al., sensory cortices. Although perceptual learning, at first
2001; Crist, Li, and Gilbert, 2001)—another indica- glance, seems to rely mostly on relatively complex
tion for the importance of top-down influences. cognitive processes, the specificity of improvement
Hence, present models of perceptual learning in- for quite low-level attributes such as visual field
creasingly emphasize that learning occurs at a num- position, stimulus orientation, and the eye used for
ber of di¤erent levels of information processing and training a visual task indicates a strong and crucial
that top-down influences from ‘‘higher’’ levels play involvement of primary (visual) cortex, where neu-
a crucial role in adjusting, in a task-dependent way, rons are still partly monocularly activated. However,
the processing of the ‘‘lower’’ levels. Indeed, recent monocular neurons in V1 usually lack orientation
experiments have provided direct evidence for strong specificity (Hubel and Wiesel 1977), hence specificity
top-down influences caused by error feedback and cannot be explained in a strictly feedforward system.
attentional processes on perceptual learning (Ahissar Dependence on attention, error feedback, and ‘‘in-
and Hochstein 1993; Weiss, Edelman and Fahle sight,’’ on the other hand, demonstrates that strong
1993; Herzog and Fahle 1997; see also chapters 18 top-down influences play a major role in perceptual
and 20). Di¤erent observers may use (slightly) dif- learning and that perceptual learning, of course, also
ferent strategies to solve the same task, possibly involves more cognitive levels of the brain. Hence,
involving di¤erent levels of cortical stimulus repre- the study of perceptual learning processes not only
sentation, and may improve at di¤erent speeds. These shows us the amazing amount of plasticity even in
di¤erences may cause one of the unsolved problems adult sensory information processing at a relatively
in perceptual learning: Why is it that some observers peripheral level but also leads to a view of cortical
improve vigorously while others not at all? information processing not as a feedforward system
It seems that the high specificity of perceptual of sequential neuronal layers but as a complex and
learning in vision is partly lost if relatively easy tasks plastic feedback system with strong and important
are learned, while specificity of improvement is top-down influences on ‘‘lower’’ or ‘‘early’’ parts of
highest for very di‰cult tasks. A possible explanation information processing.
is that ‘‘easy’’ tasks are learned on a relatively ‘‘higher’’
level of information processing, where the informa-
Recent Advances in Research on Perceptual
tion extracted from the visual scene is better used
Learning
than was possible before training and hence show
interocular transfer of improvement. Di‰cult tasks,
Perceptual learning is a quite active area of research,
on the other hand, may require additional changes
and the extensive list of references at the end of
on lower levels of processing that are specific for the
Introduction xv
this book is testimony to this activity. Some of this 1989; Hughes, Epstein, Schneider, and Dudock
activity is outside the scope of the individual chapters 1990), but also with other tasks such as tactile re-
of this book. On the following pages, I will briefly versal and oddity learning (Krekling, Tellevik, and
referee important contributions to perceptual learn- Nordvik 1989; Benedetti 1991; cf. also Layton
ing, most of which are not cited elsewhere in the 1972). This tactile learning is highly task specific
book. The fast progress in research on perceptual while generalizing between di¤erent body locations
learning is documented by quite a number of reviews (Sathian and Zangaladze 1997; cf. also Nagarajan,
highlighting di¤erent aspects of this research (Gibson Blake, Wright, Byl, and Merzenich 1998; Sathian
1963; Wohlwill 1966; Hall 1991; Sagi and Tanne and Zangaladze 1998; Harris and Diamond 2000; cf.,
1994; Gilbert 1994; Walsh and Both 1997; Gold- however, Godde, Stau¤enberg, Spengler, and Dinse
stone 1998; Wiggs and Martin 1998; Hurlbert 2000). 2000). These psychophysical studies are in agreement
Perceptual learning has a number of di¤erent aspects, with the plasticity of the somatosensory system found
ranging from sensorimotor (e.g., tactile) to purely in both sum potential recordings (Pascual-Leone and
sensory learning in all sensory submodalities such as Torres 1993; Spengler, Roberts, Poeppel, Byl, Wang,
olfaction, taste, hearing, and vision. Rowley, and Merzenich 1997) and recordings from
The interplay between sensory input and motor single neurons and small clusters of neurons (Recan-
output was tested, for example, in maze discrimi- zone, Merzenich, Jenkins, Grajski, and Dinse 1992;
nations in animals (Trobalon, Sansa, Chamizo, and Wang, Merzenich, Sameshima, and Jenkins 1995;
Mackintosh 1991; Trobalon, Chamizo, and Mac- Diamond, Amstrong-James, and Ebner 1993; Godde,
kintosh 1992; cf. also Prados, Chamize and Mac- Spengler, and Dinse 1996).
kintosh 1999), indicating the importance of visual Perceptual learning of taste and olfaction leads to
context on some forms of motor learning. Spatial bias-free improvement in wine discrimination (Owen
learning, as tested with a prism adaptation task, and Machamer 1979; Bende and Nordin 1997; for a
revealed the existence of di¤erent levels of learning. review, see Granger and Lynch 1991). Generalization
More complex spatial remapping can be learned on of a conditioned aversion reaction to lemon-saline
a cognitive level but not on a perceptual level to presentation of lemon-sucrose was substantially
(Bedford 1993; cf. also Howard 1971). The cerebel- reduced by prior exposure to the individual com-
lum may play a role in this type of learning (cf. pound solutions (Mackintosh, Kaye, and Bennett
Daum et al. 1993; Vaina, Belliveau, Des Roziers, 1991; cf. also Bennett and Mackintosh 1999; Bennett,
and Ze‰ro 1998). Several studies proved the intri- Levitt, and Anton 1972; Bennett and Anton 1992;
cate relation between visual perception and motor Anton, Player, and Bennett 1981 for the e¤ects of
learning in which demonstrations of a motion stimulus preexposure on perceptual learning in rats).
allowed human observers to adequately reproduce Auditory learning concerns either more basal or
this motion (Vogt 1995; cf. also Meegan, Aslin and else complex functions, such as understanding of
Jacobs 2000; cf. Kodman 1981 for visual-motor language. On a basal level, training to discriminate
learning in retarded persons). between di¤erent frequencies increases precision of
Distinct perceptual learning occurs in the somato- performance especially for the trained frequency band
sensory system, for example with vibrotactile stimu- in humans (Demany 1985; Irvine, Martin, Klimkeit,
lation (Epstein, Hughes, Schneider, and Bach-y-Rita and Smith 2000) as well as the cortical representation
xvi Introduction
of this frequency band in the monkey auditory cortex Gliner, Mihevic, and Horvath 1983). Training an
(Recanzone, Schreiner, and Merzenich 1993). Per- orientation-discrimination task led to a decrease of
formance in several auditory tasks improves through activity in striate and extrastriate visual cortex during
learning (Grunke and Pisoni 1982; Tomblin and task execution, as measured with positron emission
Quinn 1983), with complex tasks requiring longer tomography (Schiltz, Bodart, Dubois, Dejardin,
learning times (Watson 1980; cf. also Leek and Wat- Michel, Roucoux, Crommelinck, and Orban 1999),
son 1988). Comparisons between monaural and bin- while it increased activity in prefrontal cortex during
aural loudness estimates indicate the basis of learning an associative sensory learning task (McIntosh, Rajah,
during calibration of a loudness estimate (Marks, and Lobaugh 1999). Bilateral electroconvulsive
Galanter, and Baird 1995). Language comprehension therapy had a greater impact on perceptual learning
improves through training (reviewed by Kuhl 1994), than right unilateral ECT (Daniel, Crovitz, and
and training with sentences improves performance Weiner 1984).
more than training with isolated words (Greenspan, Animal models of perceptual learning on the be-
Nusbaum, and Pisoni 1988; Nygaard and Pisoni havioral level found that mixing exposure of stimuli
1998; cf. also Harpur, Estabrooks, Allen, and Asaph to be discriminated accelerated perceptual learning
1978, for perceptual learning in language acquisi- compared to separate exposure (Honey and Bateson
tion). Improvement may last for a long time (Bra- 1996; cf. also Honey, Bateson, and Horn 1994).
dlow, Akahane-Yamada, Pisoni, and Tohkura 1999), Di¤erent animal studies demonstrated e¤ects of ani-
a finding important for perceptual learning with mal population density on perceptual learning (Levitt
cochlear implants (Watson 1991; Pisoni 2000; see and Bennet 1975), investigated the relationship be-
Clark, chapter 8, this volume). tween imprinting and perceptual learning in chicks
Not only in the somatosensory but also in the (Kovach, Fabricius, and Fält 1966), led to insights
visual system, receptive field properties change as a into the perceptual learning of monkeys (Hum-
result of perceptual learning (Eysel, Eyding, and phreys and Keeble 1976; Ga¤an 1996), and showed
Schweigart 1998; Gilbert 1994; Zohary, Celebrini, that early visual deprivation impairs visual discrimi-
Britten, and Newsome 1994; Logothetis, Pauls and nation learning (Zernicki 1999).
Poggio 1995; Gilbert, Das, Ito, Kapadia, and West- The mechanisms underlying perceptual learning
heimer 1996; Tovee, Rolls, and Ramachandran have been extensively investigated and discussed.
1996; Gilbert, Ito, Kapadia, and Westheimer 2000; One central question is specificity versus generaliza-
Eysel, chapter 3, this volume). Consistent with this tion: while perceptual learning is quite specific, cer-
plasticity on the single-cell level, sum-potentials tain aspects nevertheless generalize. For example, a
over the human occipital pole change after training second, similar task may start at the same low per-
of motion- and stereo-discrimination tasks (Fahle formance level as the initial task, but performance
and Skrandies 1994; Skrandies and Fahle 1994; may improve faster than in the initial task (Liu and
Skrandies 1995; Skrandies, Lang, and Jedynak 1996; Weinshall 2000). This finding is interpreted using
Skrandies and Jedynak 1999; Skrandies, Jedynak, and a model based on limited computational capacity.
Fahle 2001), as well as during associative learning Abrupt learning can be interpreted as either being
(Miltner, Braun, Arnold, Witte, and Taub 1999) and based on di¤erent mechanisms than gradual learning
sensorimotor learning (mirror-star learning task: (Ahissar and Hochstein 1997, chapter 14, this vol-
Introduction xvii
ume) or else on the same mechanisms (Nakayama more formal models of perceptual learning have been
and Shapley; Rubin, Nakayama, and Shapley 1997; proposed (see chapters 15, 18, and 20, this volume).
Rubin et al., chapter 13, this volume), but it seems Some of these are focussed on receptive field orga-
clear that di¤erent levels of cortical processing are nization (e.g., Kalarickal and Marshall 1999), while
involved (cf. also Watanabe et al., 2001). Other ex- most are concerned with more complex features
periments indicate that the target signal changes (Peres and Hochstein 1994; Rentschler, Jüttner,
during perceptual learning while the level of internal and Caelli 1994; Stone 1996; Jüttner, Caelli, and
noise stays constant (Gold, Bennett, and Sekuler Rentschler 1997; Stone and Harper 1999; Saksida
1999; cf., Dosher and Lu 1998; Dosher and Lu 1999). 1999).
Incomplete pictures are far better recognized after Improvement through training occurs for several
perceptual learning that probably takes place on a low-level tasks such as motion perception (Liu and
rather cognitive level (Gollin 1965), and one study Vaina 1998; Liu 1999; Zanker 1999), orientation
reports on the relationship between autobiographical discrimination (Rivest, Boutet, and Intriligator 1996;
(explicit) and perceptual (implicit) learning ( Jacoby King, Shanks, and Hart 1996; Fahle 1997; cf. Baldassi
and Dallas 1981). and Burr 2000 for the relation between visual search
Attention exerts a significant influence on many and orientation identification versus target location),
types of perceptual learning (e.g., Adcock and discrimination of spatial frequencies (Fiorentini and
Mangan 1970; Ahissar and Hochstein 1993; Ito, Berardi 1980, 1997) and textures (Karni and Sagi
Westheimer, and Gilbert 1998). It has been discussed 1991), as well as in seeing form from motion (Vidya-
whether long-term changes in perception should be sagar and Stuart 1993) and spatiotemporal interpola-
explained by association or perceptual learning tion (DeLuca and Fahle 1999). Visual hyperacuity
(Bedford 1995, 1997) and whether inhibitory pro- denotes the fact that discrimination thresholds in
cesses may be involved in perceptual learning (Kill- several visual tasks such as stereoscopic depth percep-
cross, Kiernan, Dwyer, and Westbrook 1998). The tion, vernier acuity, and orientation discrimination
dimensionality of internal object representations is are far below the diameter of foveal photoreceptors
reduced in peripheral vision but can be increased by and thus are sensitive probes for perceptual learning.
perceptual learning ( Jüttner and Rentschler 1996). Several studies found improvement through training
Some forms of perceptual learning may require REM with these tasks to be highly specific for most stimulus
sleep (Karni, Tanne, Rubenstein, Askenasy, and Sagi parameters such as visual field position, size, orienta-
1994; cf. also Karni and Sagi 1993). A large literature tion, and even the eye used during training (Freeman
exists on so-called prism adaptation e¤ects that com- 1966; Schoups, Vogels and Orban 1995; Crist,
pensate for the changes induced by wearing prism Kapadia, Westheimer, and Gilbert 1997; Matthews,
spectacles (cf. Epstein 1975; Melamed and Arnett Liu, Geesaman, and Qian 1999; cf., however, Beard,
1984). This type of perceptual learning seems to rely Levi, and Reich 1994). Stereoscopic depth percep-
mostly on changes in the motor system and is there- tion also dramatically improves through training
fore not dealt with in this book. (Ramachandran and Braddick 1973; McKee and
Based on these and other aspects of the mecha- Westheimer 1978; O’Toole and Kersten 1992; Sow-
nisms possibly underlying perceptual learning, and, den 1995; Sowden, Davies, Rose, and Kaye 1996;
of course, based on the experimental results, several van Ee 2001). Interobserver variance is sometimes
xviii Introduction
high in these tasks (Fahle and Henke-Fahle 1996). improvement transfers completely between eyes
Details on the specificity versus generalization in (Schoups and Orban 1996) but is specific to orienta-
these types of tasks can be found in chapters 11, 12, tion, size, and position (Ahissar and Hochstein 1995;
18, and 20. It should be noted that depending on cf. also Epstein 1967; Ahissar, Laiwald, Kozminsky,
training regime, specificity of learning may be and Hochstein 1998). These results indicate that per-
strongly decreased (Liu and Vaina 1998). ceptual learning for visual search may be less specific
The e¤ects of masking may change after training than learning of visual discriminations and may take
(Dorais and Sagi 1997; cf. also Saarinen and Levi place on another level of cortical processing
1995), as does the perception of illusory contours (Sireteanu and Rettenbach 1995, 2000; cf., how-
(Gellatly 1982; Rubin, Nakayama, and Shapley ever, Ellison and Walsh 1998; Lobley and Walsh 1998).
1997), and learning of a visual search task involving Parallel visual search achieved through percep-
motion information can be adversely a¤ected by tual learning may be fast but nevertheless e¤ortful
transcranial magnetic stimulation (TMS) (Stewart, (Leonards, Rettenbach, and Sireteanu 1998). Task-
Batelli, Walsh, and Cowey 1999). This technique specific attention seems to play an important role in
also demonstrated cortical plasticity in perceptual these learning processes (Ahissar and Hochstein
learning (Walsh, Ashbridge, and Cowey 1998). 2000; cf. also Stadler 1989 for the role of awareness
Not surprisingly, improvement through training in learning during visual search).
occurs for higher-level tasks, too. Some examples Perceptual learning is a potential hope for patients
are the improvement of length estimation in the su¤ering from di¤erent disorders of visual percep-
horizontal-vertical illusion which is faster under tion, ranging from amblyopia (Levi and Polat 1996)
feedback conditions than without feedback (Brosvic, over prosopagnosia (cf. Greve and Bauer 1990 for
Rowe-Boyer, and Diho¤ 1991), perceptual-motor implicit face learning; cf., however, Sergent and
skills in dentistry (Birch 1976), and rehabilitation of Villemure 1989) to brain damage with visual field
handicapped children (Scott 1974). Perception of defects (Kasten and Wuest 1998; Kasten, Wuest, and
more complex forms and objects such as textures, Sabel 1998; cf. Markowitsch and Härtling 1996 for
complex graphical signs, objects, and faces improves the e¤ects of visual priming), and Alzheimer’s disease
through practice (McLaren 1997; Espinet, Almaraz, (Postle, Corkin, and Growdon 1996), and Grafman,
and Torres 1999; Fine and Jacobs 2000; Furmanski Weingartner, Newhouse, Thompson, Lalonde, Lit-
and Engel 2000; cf. Dolan, Fink, Rolls, Booth, van, Molchan, and Sunderland (1990) report positive
Holmes, Frackowiak, and Friston 1997 for a func- ‘‘implicit’’ perceptual learning. The same is true for
tional neuroimaging correlate: enhanced activity in the somatosensory domain (see chapter 2).
inferior temporal cortex). Visual classification, such One study finds better discrimination between
as the sexing of young chickens (Biederman and novel stimuli in autistic patients than in normals but
Shi¤rar 1987), also improves as a result of training lack of perceptual learning for these stimuli in the
(Hock, Webb, and Cavedo 1987; Wills and McLaren autistic patients (Plaisted, O’Riordan, and Baron-
1998). Cohen 1998). Several studies agree that even amne-
Visual search, that is, the decision whether or not a sic patients can be primed in visual perception that
defined target is present in a display, is another lasts for up to 12 months (and hence should be con-
favorite task to test perceptual learning. For this task, sidered as perceptual learning) (Crovitz, Harvey, and
Introduction xix
Clanahan 1981; Gabrieli, Milberg, Keane and Cor- Therefore, the present book has many authors, each
kin 1990; Tulving, Hayman, and Macdonald 1991; presenting a part of the field. This, moreover, allows
Kapur, Abbott, Footitt, and Millar 1996; Chun and the reader to spot some more or less subtle di¤erences
Phelps 1999; cf. Yamashita 1993 for perceptual- between the interpretations of di¤erent experts in
motor learning in amnesic patients and Squire and the same field. It is no secret that there is a bit of
Zola 1997 for an overview of amnesia and memory). controversy among researchers on perceptual learn-
Age dependence of perceptual learning was ing regarding such issues as the exact model for the
addressed by a few studies (Fahle and Daum 1997: mechanisms underlying perceptual learning, and
fine-grain spatial memory; Gilbert 1996: visual the reader may become aware of a couple of smaller
search; Rockstroh, Dietrich, and Pokorny 1995: re- controversies regarding the best strategy to further
action time for visual stimuli). Generally, perfor- examine the phenomena and mechanisms of per-
mance deteriorates slightly with increasing age. A ceptual learning. On the other hand, I was surprised
number of additional studies addressed perceptual when reading all the chapters by how much agree-
leaning specifically in infants and children, both ment has been reached among di¤erent groups on all
normal ( Jensen 1966; Kerpelman 1967; Odom, the main issues (and most of the small ones).
McIntyre, and Neale 1971; Turnure 1972) and ab- The topics dealt with in the book cover the field
normal (Brodlie and Burke 1971; Zelniker and of perceptual learning in a rather broad and exten-
Oppenheimer 1976; De Filippo Lutzer 1986), with sive way, starting from the anatomical and physio-
generally positive results. logical changes that correspond on a neuronal level
to the changes of behavior observable on the systems
level. Starting on a neuroanatomical level, we look
Aims and Scope of This Book
at the development of interneuronal connections
during ontogenesis and at the changes induced by
The aim of the present book is twofold: first, to re-
lesions or external influences (chapter 1). Chapter 2
view the advances made during the last decade in the
then presents an overview of the changes induced
field of perceptual learning, presented by some of
by lesions and especially perceptual learning in the
the most prominent and competent researchers in the
somatosensory system. The electrophysiological cor-
field; second, to combine this information to allow a
relate of ( perceptual) learning as evidenced by the
clearer understanding of where we stand, which
plasticity of visual receptive fields after lesions and
questions have been answered, and which ones still
training, as well as the probable underlying mecha-
await a solution. So while the book can only illus-
nisms, are described in chapter 3, followed, in 4
trate the present state of knowledge on perceptual through 6, by the description of development of
learning, it also demonstrates the ways in which in-
complex receptive field properties through training
formation can be gathered, what the basic questions in temporal (chapter 4) and parietal as well as occi-
are—independent of the present state of knowledge, pital visual cortices (chapters 5 and 6). An account
and where we presently stand in the attempt to an- on cortical reorganization and perceptual learning
swer these fundamental questions. In a fast-changing and their perceptual correlates as evidenced by non-
field such as perceptual learning, no single person invasive methods in humans (chapter 7) finishes this
oversees all the newest results and developments. first part of the book.
xx Introduction
The second part of the book deals mostly with quite elementary stimulus attributes such as position
learning on a systems level as investigated mainly in the visual field, orientation, and for the eye used
by psychophysical methods. Chapters 8 to 12 con- during learning (if one eye is covered). This specif-
centrate on more low-level aspects of perceptual icity indicates a neuronal substrate for some forms of
learning, such as auditory learning (chapter 8), the perceptual learning partly on the level of the primary
adaptation and improvement through learning in visual cortex. Present models assume a plasticity of
discriminating simple visual stimuli such as gratings this primary visual cortex even in adults, triggered
(chapter 9), the long-lasting and fast modifications by appropriate stimuli under strong top-down influ-
of low-level visual networks (chapter 10), and the ences from more ‘‘cognitive’’ levels of the cortex.
invariance versus specificity of perceptual learning These findings promise to develop better therapies
(chapters 11 and 12). The chapters in part III deal for patients su¤ering from cortical lesions after stroke
with higher-level perceptual learning and address the (infarction) or trauma, based on the insights gained
relations between cognitive aspects of learning and by studying the mechanisms of perceptual learning.
perceptual learning (chapter 13) and summarize the
evidence for a bidirectional relationship between Manfred Fahle
perceptual learning and cognition (chapter 14), fol-
lowed by accounts on higher learning of early visual
tasks (chapter 15) and learning to recognize objects
in general (chapter 16), or specifically to recognize
faces (chapter 17).
The three chapters of part IV finally put it all al-
together, by modeling perceptual learning in general
(chapter 18), providing an independent component
analysis for perception (chapter 19), and by listing
some problems encountered in modeling perceptual
learning that command the development of ‘‘feed-
back’’ models of learning (chapter 20).
The glossary defines some of the central terms of
the book, and a single reference list for the entire
book will help the reader get a rapid overview. For
the hurried reader, the take-home message of the
book is given here, in order to save him or her from
reading the entire book:
Perceptual learning is a new field in the general
domain of ‘‘learning.’’ More specifically, it is a form
of implicit learning and denotes a lasting modification
of behavior following sensory stimulation caused by
previous experience with a perceptual task. In many
instances, perceptual learning is very specific for
Anatomy and Physiology
I
This page intentionally left blank
Experience-Dependent Plasticity of
Intracortical Connections
Siegrid Löwel and Wolf Singer
1
Abstract
1.1 Introduction
One of the basic features of visual scene analysis is to
assemble the components of objects into a unified percept One of the basic functions of the cerebral cortex is
and to segregate them from background. In addition,
the representation and analysis of relations among
we store information about previous experience partly
through perceptual learning and test our interpretations of components of sensory and motor patterns. To cor-
the visual world against incoming sensory input. In recent rectly interpret visual scenes, for example, the cortex
years, evidence has accumulated indicating that long-range must, on the one hand, assemble components of
neuronal connections within visual cortex mediate the an object into a unified percept and, on the other,
influences of context and experience, possibly also those of segregate these components from the background.
expectation. Focusing on the ‘‘hardware’’ of these impor-
Neurons in the primary visual cortex fire action
tant computations, namely on the layout and plasticity of
long-range connections in young and adult cortex, this potentials in response not only to the appearance of a
chapter addresses the following questions. What is the ex- particular stimulus within their classical receptive
tent and laminar specificity of these long-range con- field but also to more global characteristics of a visual
nections? What does their layout look like? Which types of scene, such as the contours and surfaces within which
neurons are involved? What is the functional role of the
a stimulus is embedded. Contrary to long-standing
connections? How do they develop? How does experience
modify their layout? Is plasticity also possible in adult belief, recent physiological and anatomical evidence
brains? What are the underlying mechanisms? suggests that spatial integration occurs in part already
at the level of the primary visual cortex. The ana-
tomical substrate for integrative capabilities at this
level and for their possible experience-dependent
Acknowledgments
modifications are long-range tangential connections
formed by excitatory cortical neurons. Because these
It is with pleasure that we thank Charles Gilbert,
horizontal connections span a cortical region much
Kevan Martin, Bill Bosking, and Rainer Goebel for
larger than that corresponding to the classical recep-
providing us with original versions of their published
tive field of an individual neuron, the connections
figures; Renate Ruhl and Ste‰ Bachmann for help
are thought to be important for context-dependent
with the figures; John M. Crook and Sven Meyburg
modifications of neuronal responsivity and thus to
for critical reading of the manuscript; and the Max-
be essential for the integration of information from
Planck-Society and the Wissenschaftsgemeinschaft
widely distant points in the visual field. And because
Gottfried Wilhelm Leibniz for their support.
4 Siegrid Löwel and Wolf Singer
the architecture of long-range connections has been 1.2.1 Extent and Laminar Specificity of Long-
shown to be modifiable by visual experience during Range Connections
early development, it follows that the criteria for
perceptual integration or grouping are at least partly The classical view of cortical connectivity, derived
acquired through experience and learning. In addi- mainly from the impregnation of single neurons
tion, in the adult, the e¤ectiveness of intracortical with silver salts (the so-called Golgi technique used
interactions continues to be modifiable on timescales in the classical studies of Santiago Ramón y Cajal and
ranging from seconds to months, and long-range hor- Camillo Golgi), is that axons tend to run perpen-
izontal connections have been implicated in these dicularly to the cortical surface, from layer to layer,
processes, too. with relatively little spread parallel to the cortical
We will first focus on the layout of long-range surface (e.g., Lund 1973). Findings of this kind sup-
intracortical connections, then discuss experience- ported the famous columnar concept of the visual
dependent modifications observed during early post- cortex, which was based on the classical electro-
natal development, and finally summarize available physiological experiments of Hubel and Wiesel
evidence for the role of these connections in adult (1962), and which particularly emphasizes vertical
plasticity. Because most of the relevant published connections (see also Gilbert 1983). However, more
data are from the mammalian visual system, we than twenty-five years ago, it was demonstrated that
will concentrate on long-range connections in the long-range connections are also a prominent feature
primary visual cortex, also termed area 17 by the of visual cortical connectivity: after cortical lesions
anatomist Korbinian Brodmann (1909), who divided in macaque monkeys, degenerating terminals were
the human cortex into consecutively numbered cyto- observed up to a distance of 3.5 mm from the
architectonic areas. lesion sites (Fisken, Garey, and Powell 1975; see also
Szentágothai 1973; Creutzfeldt et al. 1977). Experi-
ments using more modern tracer techniques (i.e.,
1.2 Layout of Intracortical Connections
extra- and intracellular injections of the enzyme
horseradish peroxidase, HRP; injections of biocytin
A fundamental characteristic of the cerebral cortex is
or fluorescent latex microspheres) have clearly estab-
the similarity of its gross morphological organization
lished that extensive horizontal connections do span
across di¤erent areas. In all mammals, the neocortex
distances of several millimeters (up to 8 mm) within
consists of six layers extending over about 2 mm
individual cortical layers (Rockland and Lund 1982;
from layer I, which lies next to the pial surface of the
Gilbert and Wiesel 1983; Martin and Whitteridge
brain, down to the lower margin of layer VI, which 1984; Kisvárday and Eysel 1992; Luhmann, Martı́nez-
directly faces the underlying white matter, a more or
Millán, and Singer 1990; for a review, see Gilbert
less cell-free region containing axons both entering
1992). Called ‘‘long-range horizontal,’’ ‘‘tangential,’’
and leaving the cortex. The cortex itself is referred to
or ‘‘intralaminar,’’ these connections are especially
as ‘‘gray matter’’ and is composed of neuronal cell
prominent in cortical layers II/III and V; their syn-
bodies, dendrites, axon terminals, and glial cells (for
apses exhibit the morphology of excitatory synapses
a general introduction to neuroscience see, for ex-
and contact dendrites of spiny and nonspiny cells in
ample, Purves et al. 1997).
the same proportion that these cell types occur in the
Experience-Dependent Plasticity of Intracortical Connections 5
cortex (80% and 20%, respectively; Kisvárday et al. (Kisvárday et al. 1997; Crook, Kisvárday, and Eysel
1986; McGuire et al. 1991). 1998). Because we are interested in long-range con-
nections that are involved in integrating visual in-
1.2.2 Types of Neurons Involved in Long-Range formation across columnar boundaries from distant
Connections points in the visual field, we will concentrate on
excitatory long-range connections.
Axon collaterals of spiny stellate and pyramidal
neurons, the two types of cortical excitatory cells 1.2.3 Patchy Nature of Long-Range Connections
(Gilbert 1983), form the anatomical basis for long-
range intracortical connections. Pyramidal cells con- Intracellular injections of HRP reveal that the axon
stitute the main cell type in the visual cortex and collaterals of individual neurons are not distributed
their somata are distributed in all cortical layers, ex- homogeneously across the cortex but—as viewed
cept layer I. They are characterized by a triangular from the cortical surface—form numerous discrete
cell body and a large apical dendrite directed radially terminal clusters. These collateral clusters measure
toward the pial surface. Typically, all their dendrites about 300 mm–600 mm in diameter. Extracellular
(apical as well as basal ones) are covered with a high injections of a variety of other tracers result in a
density of spines, dendritic protrusions that receive at similar picture of patchy intracortical connections
least one excitatory synapse (Peters and Kaiserman- (figure 1.1). These patchy connections are recipro-
Abramof 1969). Spiny stellate cells lack a dominant cal: small intracortical injections of both anterograde
apical dendrite but—as the name indicates—have and retrograde neuronal tracers label clusters of
spiny dendrites. In contrast to pyramidal neurons, (retrogradely labeled) neurons and (anterogradely
however, spiny stellate cells are present exclusively labeled) axon terminals that are spaced at regular in-
in layer IV of primary sensory areas (Lund 1973; tervals (about 1 mm) in the visual cortices of cats,
Valverde 1986). Therefore they are the major target ferrets, tree shrews, and primates (Rockland and
neurons of thalamocortical a¤erents (in case of visual Lund 1982, 1983; Rockland 1985; Gilbert and Wiesel
cortex, a¤erents from the lateral geniculate nucleus, 1983; Boyd and Matsubara 1991; Kisvárday and
a thalamic nucleus that receives information from Eysel 1992; Burkhalter, Bernardo, and Charles 1993;
the retinas) that terminate in the cortical input layer Livingstone and Hubel 1984). Thus long-range tan-
IV (Gilbert 1983). gential axon collaterals interconnect regularly spaced
About 20% of the total number of cortical neu- clusters of cells.
rons are immunopositive for gamma-aminobutyric
acid (GABA), the major inhibitory neurotransmitter 1.2.4 Modular Selectivity
in the cerebral cortex. These neurons are categorized
into at least eight di¤erent classes according to mor- Because neurons that respond to similar visual stimuli,
phological criteria. Among these, only large basket for example, to lines of a particular orientation pre-
and dendrite-targeting cells provide lateral inhibitory sented at a particular location in the visual field
connections extending up to 1.5 mm; all other cell (within their receptive field), are not distributed
types have predominantly local axon collaterals. randomly across the cortex but are arranged in col-
Thus the inhibitory network is at least two to three umns extending from layer I to layer VI (the so-
times smaller in extent than the excitatory network called orientation columns; Hubel and Wiesel 1962;
6 Siegrid Löwel and Wolf Singer
Figure 1.1
Patchiness of long-range intracortical connections in the cerebral cortex. (A) Camera lucida reconstruction of a spiny stellate
cell of cortical layer 4 displaying an extensive and patchy axonal distribution. The cell body of this neuron is located
in layer 4A, whereas most of the collateral branches are restricted to cortical layers 2 and 3. Frontal section through the
primary visual cortex of a cat. (Modified from Martin and Whitteridge 1984.) (B) Reconstruction of the axon arbors of a
pyramidal cell of layer 2 projected onto a plane parallel to the cortical surface. The cell body is located in the center of the
reconstruction. Again the clustered nature of the collateral branches is clearly visible. (Modified from Gilbert and Wiesel
1983, 1120, figure 2b.) In panels A and B, neurons were intracellularly injected with the enzyme horseradish peroxidase. (C,
D) Patchy intracortical connections visualized after extracellular injections of the fluorescent carbocyanine dye DiI. Fluo-
rescent photographs of a section cut perpendicularly to the cortical lamination ( panel C) and a section cut parallel to the cor-
tical surface ( panel D) of cat visual cortex. The dye DiI labels both axon terminals (visible as the more homogeneous white
labeling) and neuronal cell bodies (the small brightly fluorescent dots seen in panel C). Scale bar: 500 mm ( panels C, D).
Experience-Dependent Plasticity of Intracortical Connections 7
Figure 1.2
Columnar specificity of long-range tangential connections. Schematic drawing of the topographic relationship between in-
trinsic connections and orientation columns in the primary visual cortex of cats. (A) Pattern of orientation columns (black
regions represent cortical regions stimulated with horizontal moving contours). (B) Patchy distribution of retrogradely
labeled neurons (gray dots) in the same region of cortex (labeled from the injection site, marked with a star). (C) Super-
position of panels A and B. Note that the injection site of the fluorescent neuronal tracer was located in a black column (i.e.,
a horizontal orientation column) and that labeled neurons are predominantly but not exclusively distributed within columns
of the same functional preference (other dark columns).
Blasdel and Salama 1986), the question arises, do inter- of long-range connections shows a clear relationship
connected cells share similar functional properties? to the system of orientation columns: injections of
Within the last ten to fifteen years, evidence has fluorescent latex microspheres into, for example, a
accumulated that this is indeed the case with respect horizontal orientation column revealed retrogradely
to preference for stimulus orientation or color (Ts’o, labeled neurons that were concentrated in the same
Gilbert, and Wiesel 1986; Ts’o and Gilbert 1988; and other horizontal orientation columns (figure
Gilbert and Wiesel 1989; Gray et al. 1989; Hata et 1.2). In agreement with these data, the statistical
al. 1991; Malach et al. 1993; Malach, Tootell, and technique of cross-correlation analysis, used to ana-
Malonek 1994; Kisvárday et al. 1997; Bosking et al. lyze the timing of action potentials in pairs of neu-
1997). The relationship between horizontal con- rons and to measure the e¤ective connection
nections and cortical functional architecture was de- strength between two cells, showed that neurons
termined through cross-correlation analysis and by with correlated action potentials had similar orien-
combining the labeling of both intracortical con- tation preference. These findings do not mean that
nections and cortical columns, among other tech- horizontal fibers exclusively connect neurons with
niques. Anatomically, it was shown that the clustering identical response properties—cells with di¤ering
8 Siegrid Löwel and Wolf Singer
Figure 1.3
Axial selectivity of long-range intracortical connections. (A, B) Distribution of labeled synaptic boutons resulting from focal
extracellular injections of biocytin into functionally characterized sites of tree shrew visual cortex (area V1). The preferred
orientation at the injection site is shown in the inset of each panel. The outline of the dorsal portion of V1 is indicated by a
thin black line. Along the central portion of the border between areas V1 and V2, this line corresponds to a vertical line in
the map of visual space. The axis in cortex corresponding to the preferred orientation is indicated by the gray rectangle
underlying each distribution. Each black dot indicates an individual labeled bouton. There is a dense and fairly uniform
distribution of boutons found near each injection site. At longer distances the distributions are patchy, contacting other sites
that have similar orientation preference as the injection site. Note in addition that in both panels A and B the distributions
are elongated along an axis that corresponds to the preferred orientation of the injection site. (Modified from Bosking et al.
1997, figures 8A and 8C.)
properties are also contacted, although to a smaller that corresponds to the neuron’s preferred orienta-
extent—but rather that, on average, like tends to tion (Fitzpatrick 1996; Bosking et al. 1997; see
connect to like. also Schmidt, Goebel, et al. 1997; figure 1.3). Thus
long-range horizontal connections preferentially link
1.2.5 Axial Selectivity neurons with cooriented, coaxially aligned receptive
fields.
Interestingly, the long-range horizontal fibers do not
distribute isotropically across the visual cortex. In 1.2.6 Possible Functions of Long-Range
addition to the modular specificity, they exhibit axial Horizontal Connections
specificity, that is, there is a systematic relationship
between a neuron’s orientation preference and the These anatomical data suggest a close relation be-
distribution of its axon arbors across the cortical map tween the topology of tangential intracortical con-
of visual space: horizontal connections extend for nections and the perceptual grouping criterion of
longer distances and give rise to a larger number of colinearity. Dating back to the time of the Gestalt
terminal boutons along an axis of the visual field map psychologists at the beginning of the last century, it
Experience-Dependent Plasticity of Intracortical Connections 9
early in development, cells extend long, relatively orientation columns are thought to be selectively
unbranched axon collaterals. At this time, axonal eliminated and collaterals to ‘‘correct’’ orientation
arbors lack distinct collateral branches. Between three columns, to be added (Katz and Callaway 1992).
and six weeks postnatally, when clearly segregated
clusters of retrogradely labeled neurons are emerg-
1.4 Plasticity of Long-Range Connections
ing, the long-range axon collaterals become grouped
during Development
and distal collateral branches are elaborated, as seen
in the adult (Katz and Callaway 1992).
The emergence of well-segregated clusters of in-
Taken together, the emergence of patchy intra-
terconnected cells in a developmental period when
cortical connections from an initially random dis-
visual experience is known to profoundly influence
tribution of interconnected cells early in life can be
divided into at least two distinct phases: (1) crude cortical development (the ‘‘sensitive period’’; see, for
example, Wiesel, 1982) indicated that the specificity
cluster emergence, followed by (2) arbor refinement,
of the long-range connections might also depend on
consisting of both selective retraction of ‘‘inappro-
visual experience. In particular, the anatomical ob-
priate’’ connections and the elaboration of ‘‘appro-
servation that clusters refine by the elimination of
priate’’ axonal branches (see also Durack and Katz
one set of connections (the ‘‘inappropriate’’ ones)
1996; Ruthazer and Stryker 1996).
and the stabilization of another set of connections
(the ‘‘appropriate’’ ones) raised the possibility that se-
1.3.1 Patchy Connections and Orientation
lective stabilization is influenced by neuronal activity
Columns during Development
and is not genetically determined.
Because of the close relationship between clustered
long-range connections and orientation columns, 1.4.1 Development of Patches
neurons located between the crude clusters in early
Experiments with visually deprived animals con-
postnatal development must receive inputs from
firmed this hypothesis. In cats that were dark-reared
di¤erent orientation columns. This is likely to ac-
or binocularly deprived (lid-sutured), the selectivity
count for the observation that orientation tuning is
of long-range intracortical connections was severely
much broader in young cats than in adults: adult
reduced, and the ‘‘normal’’ and selective adult pat-
values for tuning emerge at the same time clusters
tern of connectivity did not appear (Katz and Call-
achieve their adult level of refinement (about one
away 1992; Luhmann, Martı́nez-Millán, and Singer
month postnatally; Imbert and Buisseret 1975; Albus
and Wolf 1984; Braastad and Heggelund 1985). 1986; Luhmann, Singer, and Martı́nez-Millán 1990).
Nevertheless, crude clusters were present, so that the
Similarly, in ferrets, the second phase of cluster re-
initial phase of cluster refinement does not seem to
finement coincides with the emergence of mature
require patterned visual activity. This does not nec-
orientation tuning and maps (Ruthazer and Stryker
essarily indicate that neuronal activity plays no role
1996; Durack and Katz 1996). In summary, in the
in cluster development. Indeed, chronic infusion of
early phase of crude cluster emergence, neurons make
tetrodotoxin (TTX, a sodium channel blocker that
connections with a larger-than-normal range of ori-
prevents neurons from firing action potentials)
entation columns. Thereafter, when adultlike clus-
into the visual cortex of ferrets resulted in a spatially
ters appear, axon collaterals projecting to ‘‘incorrect’’
12 Siegrid Löwel and Wolf Singer
random distribution of retrogradely labeled cells, ened, whereas synaptic contacts between asynchro-
whereas removal of the eyes did not prevent the nously active pre- and postsynaptic neurons will be
initial development of crude clusters (Ruthazer and weakened. Correlation-based mechanisms inspired
Stryker 1996). Thus blockade of cortical but not of by Hebb’s original ideas about the modification of
retinal activity prevented the initial development of synapses have been implicated in a variety of sys-
clustered horizontal connections. These observations tems, including the development of ocular domi-
indicate that spontaneous activity patterns in the nance columns in the visual cortex (Hubel and Wiesel
primary visual cortex, the lateral geniculate nucleus 1965) and in the development of the retinotectal
(its main a¤erent nucleus), or other a¤erent sources projection in goldfish and frogs (Constantine-Paton,
are su‰cient to organize the emergence of crude Cline, and Debski 1990).
clusters. Direct evidence for the hypothesis that long-range
The refinement of clusters, however, is clearly horizontal connections are also stabilized selectively
dependent on visually driven patterned activity. between cells exhibiting correlated activity was ob-
Both dark-rearing and prolonged binocular depriva- tained in strabismic cats. In these animals, the op-
tion prevent the normal progression from ‘‘crude’’ tical axes of the two eyes are not aligned, and
to ‘‘refined’’ clusters. Experiments using retrograde the images on the two retinas cannot be brought
tracers injected into area 17 revealed patterns of ret- into register. As a result, the responses mediated by
rogradely labeled neurons in binocularly deprived anatomically corresponding retinal loci in the two
cats at PND 38 that very much resembled patterns eyes are not correlated. During a critical period of
obtained in normally raised animals at PND 14 (Katz postnatal development, the connections between the
and Callaway 1992), although intracellular injec- a¤erents from the two eyes and their common cor-
tions showed that individual ‘‘deprived’’ axon arbors tical target cells are malleable and become destabi-
resembled normal arbors of similar age, displaying lized if their activity is not su‰ciently correlated
both clustered and elaborate distal branches. Thus (Singer 1990; Wiesel 1982; Swindale 1981; Miller,
neurons are likely to restructure their axonal arbors Keller, and Stryker 1989). As a consequence, stra-
but to connect to a larger than normal range of ori- bismus, also called ‘‘squint,’’ accentuates the segre-
entation columns. gation of the geniculocortical a¤erents from the two
Taken together, the development of patchy long- eyes in layer IV (Shatz, Lindström, and Wiesel 1977;
range intracortical connections is characterized by Löwel 1994), and most of the cells in the visual
an early period of crude cluster formation that is cortex become responsive exclusively to stimulation
activity- but not experience-dependent and a later of either the left or the right eye (Hubel and Wiesel
period that is experience dependent and causes cluster 1965). Although each of these monocularly driven
refinement. cell populations is capable of subserving normal pat-
What are the mechanisms underlying this experi- tern vision, strabismics are unable to combine infor-
ence-dependent stabilization of intracortical connec- mation coming from the two eyes into a single
tions? According to Hebb’s postulate (1949) for percept; to avoid double vision, they use only one
associative learning and its modern extension by eye at a time and suppress the signals from the other
Stent (1973) and Changeux and Danchin (1976), eye (Duke-Elder and Wybar 1973; von Noorden
synaptic contacts between synchronously active pre- 1990). Thus, the coherence of responses to visual
and postsynaptic neurons are selectively strength- patterns is likely to be the same in strabismics as in
Experience-Dependent Plasticity of Intracortical Connections 13
normal animals for cells driven from the same eye, opment of thalamocortical connections (see, for
but much lower for cells driven from di¤erent eyes. example, Wiesel 1982): neurons ‘‘wire together if
Anatomical experiments in the primary visual cortex they fire together.’’
of divergently squinting cats (cats with a divergent In summary, these anatomical results are all com-
squint angle use alternating fixation to avoid double patible with the idea of a selective stabilization of
vision) revealed that cell clusters were driven almost tangential fibers between coactive groups of neu-
exclusively from either the right or the left eye and rons. They support the hypothesis that the strength
that tangential intracortical fibers preferentially con- of intrinsic connections in the primary visual cortex
nected cell groups activated by the same eye (Löwel reflects the frequency of previous correlated activa-
and Singer 1992). After injections of retrograde tion. The experimental evidence appears convincing
tracers into the primary visual cortex, retrogradely for the ‘‘ocular dominance selectivity’’ of the tan-
labeled neurons were distributed in well-segregated gential fibers: these fibers tend to connect cell groups
clusters up to 5 mm from the injection site. The activated by the same eye in strabismic but not in
locations of cell groups preferentially activated by normally sighted cats. The role of visual experience
either the right or the left eye (the so-called ocular in fine-tuning the ‘‘orientation selectivity’’ and the
dominance columns; Hubel and Wiesel 1962) were ‘‘axial selectivity’’ of the tangential fibers is less clear.
visualized using the radioactive tracer [ 14 C]2-deoxy- Experiments in strabismic cats indicate that strabis-
glucose (2-DG; Sokolo¤ et al. 1977): after monoc- mus does not interfere with the tendency of long-
ular visual stimulation, regions of increased neuronal range horizontal fibers to link predominantly neurons
activity take up more of the radioactively labeled of similar orientation preference (Schmidt, Kim, et
glucose analog than less active regions. Because they al. 1997).
accumulate the radioligand, they can be visualized Recently, modular and axial selectivity of tan-
autoradiographically: regions of increased neuronal gential fibers was reported in the visual cortex of
activity and thus of increased radioactivity show up very young tree shrews as early as one to three weeks
as darker on X-ray films exposed to brain sections. after eye opening (Crowley et al. 1996). It thus
Comparing the patterns of retrogradely labeled seems that either a small amount of visual experience
neurons to those of ocular dominance columns is already su‰cient or that visual experience is not
labeled with 2-DG revealed that cell clusters were needed at all for establishing the modular and axial
located preferentially within the same ocular domi- specificity of horizontal connections as seen in adult
nance territories as the injection site (figure 1.6). animals, although further experiments are needed to
Analyses of normally sighted control animals pro- clarify this issue.
vided no evidence for an eye-specific selectivity of
tangential connections. This agreed with other evi-
1.5 Development and Plasticity of Callosal
dence that in normally sighted cats, tangential con-
Connections
nections are related to orientation but not to ocular
dominance columns (Gilbert and Wiesel 1989;
In the visual system, the left visual hemifield is rep-
Schmidt, Kim, et al. 1997). These results suggested
resented in the right visual cortex and the right visual
that the development of long-range intracortical
hemifield is represented in the left visual cortex.
connections depended on experience-dependent
Nevertheless, everyday experience tells us that we do
selection mechanisms similar to those in the devel-
Figure 1.6
Experience-dependent selection of long-range intracortical connections. Topographic relations between ocular dominance
columns and intrinsic connections in the primary visual cortex of strabismic ( panels A–C) and normally raised cats ( panel D).
(A) Distribution of retrogradely labeled cells after an injection with fluorescent latex microspheres. White dots indicate the
position of individual cells; a star, that of injection site. (B) 2-deoxyglucose pattern showing the topography of ocular
dominance territories in the region containing the retrogradely labeled cells in panel A. The black regions represent the
domains of the right eye. (C) Superposition of panels A and B. Most of the retrogradely labeled cells are located within zones
of high 2-DG uptake (black regions). The injection site was located in a right eye domain as verified by the location of
retrogradely labeled cells in the lateral geniculate nucleus. (D) Superposition of ocular dominance domains (black regions)
and retrogradely labeled neurons (white dots) in normally raised cats. Note the absence of a systematic topographic rela-
tionship between the two patterns. (Modified from Löwel and Singer 1992.)
Experience-Dependent Plasticity of Intracortical Connections 15
not see a break at the connection between left and tially imprecise and exuberant and attain their adult
right visual hemifield, the vertical meridian. How is specificity by elimination of ectopic axon terminals
this achieved? The two hemispheres of the brain are (Innocenti and Caminiti 1980). As with horizontal
connected by large axon bundles, one of which is fibers, visual callosal connections can be modified
the corpus callosum. It is by far the largest fiber tract by manipulating early visual experience: early stra-
in the brain and consists almost exclusively of axons bismus, monocular deprivation, and short periods
linking cortical neurons. The main function of the of binocular deprivation interfere with the develop-
callosal fibers is to ensure that the two hemifields are mental process that eventually confines the callosal
combined into a single percept of the outside world projection zone to the vertical meridian, so that callo-
without a noticeable break at the vertical meridian. sally projecting neurons come to occupy a broader
Callosal fibers share a number of features with strip along the border between areas 17 and 18
long-range intracortical connections. They originate (Innocenti and Frost 1979; Berman and Payne 1983;
from and terminate on similar classes of cells, are re- Elberger, Smith, and White 1983; Lund, Mitchell,
ciprocal, and exhibit topological specificity (for a re- and Henry 1978; Cynader, Lepore, and Guillemot
view, see Innocenti 1986). In the adult, callosally 1981). Thus both tangential intrinsic fibers and cal-
projecting neurons are restricted to a strip (a few losal connections exhibit a high degree of selectivity
millimeters wide) at the cortical representation of the in the adult, and both projections are susceptible
vertical meridian (the border between areas 17 and to experience-dependent modifications. Studying
18). Neurons located close to the representation of callosal fibers in the primary visual cortex of strabis-
the vertical meridian have receptive fields that cross mic cats, it was shown that the selection mechanisms
the midline of the visual field, and the responses to for stabilizing callosal connections are similar to those
stimuli in the ipsilateral hemifield are conveyed by responsible for specifying the tangential intracortical
callosal input. The ipsilateral and contralateral halves connections: comparing the pattern of retrogradely
of the crossing receptive fields have the same ori- labeled neurons to that of domains sharing the same
entation and direction preference (Berlucchi and ocular dominance and orientation preference labeled
Rizzolatti 1968; Lepore and Guillemot 1982; Blake- by the radioactive 2-DG technique revealed that 60%
more et al. 1983), suggesting that the orientation of the transcallosally labeled neurons were located
preference of the callosal a¤erents is matched with in the monocular orientation domains labeled with
that of their respective target cells. Callosally pro- 2-DG (Schmidt, Kim, et al. 1997). Callosal fibers
jecting neurons in sensory, motor, and association thus interconnect neurons exhibiting the same ocu-
areas of many species exhibit a columnar distribution lar dominance and orientation preference, with sim-
(Innocenti 1986), and the axon arbors of callosal ilar selectivity as the intrinsic fibers, so that both fiber
neurons are patchy (Houzel, Milleret, and Innocenti systems seem to be equally susceptible to the e¤ects
1994). of strabismus. The development of the interhemi-
In young kittens, callosally projecting neurons are spheric pathway and that of long-range intracortical
not confined to the representation of the vertical connections are therefore likely to be governed by
meridian but distribute across the entire area 17 similar experience-dependent organizing principles;
(Innocenti, Fiore, and Caminiti 1977; Innocenti and neurons exhibiting decorrelated activation patterns
Caminiti 1980). Thus callosal axons in cats are ini- lose their corticocortical connections (Schmidt, Kim,
16 Siegrid Löwel and Wolf Singer
et al. 1997). This observation agrees with previous before the lesion responded only to the skin surfaces
suggestions that interhemispheric and intrahemi- on the digit to be amputated) now responded to
spheric connections serve similar functions and stimulation of the adjacent, intact fingers (Merzenich
should therefore be organized similarly (Hubel and et al. 1983a,b, 1984). Thus the cortical representa-
Wiesel 1967; Innocenti 1986). tion of the intact fingers changed and expanded in
size at the expense of the amputated digit’s repre-
sentation. Although this expansion was significant,
1.6 Plasticity of Long-Range Intracortical
representational changes remained relatively local
Connections in the Adult
and applied only to the representation of intact, im-
mediately adjacent digits.
Until recently, it was widely assumed that the ‘‘wir-
Further experiments demonstrated that represen-
ing’’ of the mature cortex is stable. Plasticity in the tational plasticity in the somatosensory cortex did
adult cerebral cortex was thus largely ignored as a
not necessarily require amputation. When adult owl
scientific topic, in contrast to the ‘‘rewiring’’ of
monkeys were trained to use a particular digit for a
neuronal circuitry in developing systems or phe-
behavioral task that was repeated several thousand
nomena such as long-term potentiation (LTP) and
times, the cortical representation of that digit
long-term depression (LTD) in the cortex, hippo-
expanded at the cost of the other, less often used
campus and cerebellum (see Eysel, chapter 3.6.1, and
digits ( Jenkins et al. 1990). Thus practice alone was
Sterr et al., chapter 7.1, this volume). This has
su‰cient to enlarge the region of cortex containing
changed dramatically since the pioneering experi-
neurons that were activated during the repetitive
ments of Michael Merzenich, Jon Kaas, and their behavioral task.
colleagues, who demonstrated that receptive fields
In the following years, a large body of evidence
in the adult somatosensory system can change both accumulated indicating that in adult cortex of all
their size and arrangement within sensory maps in
sensory modalities and in the motor system, neuronal
response to peripheral lesions (for a review, see
response properties and the functional architecture
Buonomano and Merzenich 1998a; Dinse and Mer-
can display experience-dependent changes in the
zenich, chapter 2, and Sterr et al., chapter 7, this vol-
fine-grain representation of the sensory surface (for a
ume). In the somatosensory system, the body surface
review, see Dinse and Merzenich, chapter 2, and
is mapped topographically onto the primary somatic
Eysel, chapter 3, this volume). In the visual system,
sensory cortex, much as in the visual system, where
map reorganization was demonstrated with retinal
the visual field is represented in the primary visual
lesions in the primary visual cortex of both cats and
cortex. In a series of experiments Merzenich and monkeys. Because most visual cortical neurons have
colleagues examined the cortical representations of
binocular receptive fields, it was necessary to lesion
the hand in owl monkeys before and after amputa- both retinas to e¤ectively deprive the visual cortex
tion of a digit (Merzenich et al. 1984). Two to eight of all its normal input. Immediately after the lesion,
months after this peripheral manipulation, the body the cortical region receiving input from the lesioned
map in the somatosensory cortex was changed sub- part of the retina was silenced, although it recovered
stantially such that neurons within the cortical area over the next few months. Within days or weeks,
originally occupied by the missing finger (cortex that cells in the initially silent cortical region (the cortical
Experience-Dependent Plasticity of Intracortical Connections 17
‘‘scotoma’’) began to fire action potentials when ret- however: areas of the lateral geniculate nucleus cor-
inal loci outside the lesion were stimulated. Thus responding to the cortical scotoma remained silent
receptive fields of cells in the cortical scotoma shifted two months after lesioning (Gilbert and Wiesel
from representing the lesioned part of the retina to 1992). In addition, conditioning experiments in the
retinal zones surrounding the lesion. As in the cortex produced clear evidence that plastic changes
somatosensory system, the representation of the can occur in adult cerebral cortex (Greuel, Luh-
lesioned area decreased, whereas the representation mann, and Singer 1988; Frégnac et al. 1992; Cruik-
of the areas around the lesion increased (Kaas et al. shank and Weinberger 1996b). Because cortical
1990; Chino et al. 1992; Gilbert and Wiesel 1992; reorganization was observed to extend over more
Heinen and Skavenski 1991, see also Schmid et al. than 10 mm laterally (Pons et al. 1991), the long-
1996; for reviews, see Kaas 1991; Buonomano and range horizontal connections were considered the
Merzenich 1998a; Gilbert 1998). In a related exper- most likely candidates to mediate these changes. In
iment, monkeys were fitted with prisms that re- the visual system, the spread of thalamocortical
versed the visual field. After a few months, neurons a¤erents is roughly 2 mm and thus much smaller
in the primary visual cortex developed novel recep- than the extent of the observed plastic changes in the
tive fields in the ipsilateral hemifield that normally cortex. In a series of experiments aimed at elucidat-
only activates neurons in the contralateral cortex ing the respective roles of cortical versus subcortical
(Sugita 1996). These results indicated that (1) visual sites for cortical reorganization, electrophysiological
cortical neurons can acquire new inputs not only recordings were performed at various stages along
from neighboring retinal areas but also from distant the visual pathway. Even after cortical reorganization
areas; and that (2) changes in the visual input—not was completed, a large and silent region approxi-
necessarily lesions—are su‰cient to induce changes mating the size of the normal representation of the
in the visual field map. lesioned retinal area remained in the lateral genicu-
late nucleus, indicating that most of the reorganiza-
1.6.1 Underlying Mechanism and Substrate tion must be intrinsic to the cortex (Gilbert and
Wiesel 1992; Darian-Smith and Gilbert 1995). Ana-
One of the major questions raised by these inves- tomical analysis in these animals demonstrated that
tigations was where in the nervous system the corti- the thalamocortical a¤erents did not sprout into the
cally recorded changes actually take place. Although cortical scotoma, so that their physical extension
it is now clear that virtually every level of the ner- (1.5–2.0 mm laterally in cortex) was insu‰cient to
vous system is able to exhibit plasticity under certain account for the observed cortical reorganization.
circumstances, visual system experiments indicated In contrast, long-range horizontal connections dis-
that in many cases the primary site of change is played sprouting. Using extracellular injections of
located in the cortex itself. In principle, both retina the anterograde tracer biocytin into cortex just out-
and the lateral geniculate nucleus could be sites of side the boundary of the original cortical scotoma,
plasticity. If this were the case, one would expect Darian-Smith and Gilbert (1994, 1995) compared
that the silent zones in these structures would shrink the density of lateral projections into reorganized vi-
in parallel with the observed cortical changes (see sual cortex to that of lateral projections into normal
chapter 11.7 this volume). This was not the case, (nondeprived) visual cortex and found that axon
18 Siegrid Löwel and Wolf Singer
collaterals from cortical neurons surrounding the vi- are thought to become unmasked, to be strengthened
sual cortical scotoma predominantly branched into with time, and eventually to become the neuron’s
the deprived as opposed to the normal cortical area: major excitatory drive. Taken together, the consensus
axon fibers were always denser within reorganized from many studies is that activity-dependent repre-
than within normal cortex, with fiber densities 57– sentational plasticity arises from a combination of
88% greater in reorganized than in normal cortex. unmasking of widespread, normally subthreshold con-
Because morphological changes consisted in an en- nectivity and the formation of new functional con-
richment of fiber clusters (both axon collaterals and nections through axonal sprouting. Strengthening the
synaptic boutons were added) rather than an exten- e¤ectiveness of existing connections may involve,
sion of fibers beyond the normal level (Darian-Smith on a timescale of minutes to hours, use-dependent
and Gilbert 1994), reorganization was most probably facilitation of the e¤ectiveness of preexisting synaptic
mediated by modifications in the preexisting frame- connections, whereby the modifications may be ini-
work of long-range intracortical connections. Simi- tiated by phenomena such as LTP and LTD (Singer
larly, in the motor cortex, functional reorganization 1995; see also Eysel, chapter 3, this volume). Over
was mediated and constrained by the anatomical longer time periods (weeks to months), cortical re-
framework of preexisting, horizontal projections organization involves the formation of new synapses
(Huntley 1997). On the other hand, a recent study and the sprouting of additional axon collaterals and
(Florence, Taub, and Kaas 1998) indicated that thus the establishment of entirely new connections.
sprouting in the somatosensory system may also oc- Unexpected just a few years ago, adult plasticity
cur beyond the framework of preexisting con- thus displays properties previously thought to be
nections: analysis of the distributions of thalamic and restricted to so-called critical periods early in devel-
cortical connections in macaque monkeys with long- opment. Contrary to long-standing belief, cortical
standing accidental trauma to a forelimb revealed representations in adult animals are thus highly dy-
that thalamocortical projections were relatively nor- namic entities, continuously modified by experience
mal, whereas connections in the somatosensory cor- and adjusted to current environmental demands:
tex (areas 3b and 1) were markedly more widespread whenever a particular peripheral input is used pro-
in lesioned than in normal animals. portionally more than other inputs, more cortical
The receptive field shift of neurons in the original area seems to be allocated to the representation of
cortical scotoma indicates that the role of horizontal that input. A prime example of these plastic changes
long-range connections changes from modulating is a recent fMRI study (Karni et al. 1995) in the
a neuron’s response (subthreshold) to driving it human primary motor cortex showing a more ex-
(suprathreshold). Normally, visual cortical neurons tensive representation of a practiced versus unprac-
respond best to stimulation of a small portion of the ticed sequence of finger movements after several
visual field (their ‘‘classical’’ receptive field), but weeks of daily training sessions. Persisting for several
nevertheless receive synaptic connections from out- months, these changes suggest that a long-term ex-
side their classical receptive field. When the principal perience-dependent reorganization of the adult mo-
a¤erent input to a visual or somatosensory cortical tor cortex may underlie the acquisition and retention
neuron is removed (as in the case of retinal lesion or of a new motor skill.
digit amputation experiments), subthreshold inputs
Adaptation of Inputs in the Somatosensory
System
Hubert R. Dinse and Michael M. Merzenich
2
Abstract Because, for example, amputation changes the
pattern of use entirely, a more accurate distinction
This chapter summarizes evidence that cortical maps and would be between ‘‘lesion-induced’’ and ‘‘non-
cortical response properties are in a permanent state of lesion-induced’’ plasticity. To what extent the two
use-dependent fluctuations, where ‘‘use’’ includes training-
forms are based on di¤erent or perhaps even on
and learning-induced changes. In their simplest form, use-
dependent changes are input driven. Although attention similar mechanisms is a matter of ongoing debate.
and other high-level processes may contribute and enhance In contrast to developmental plasticity, adapta-
use-dependent neural changes by specific pathways con- tions of adult brains do not rely on maturation or
veying top-down information, reorganization can occur in growth. For learning-induced alterations, there is
the absence of high-level processes. The current experi-
agreement on the crucial role played by so-called
mental data imply that altered performance is based on
altered forms of neural representations, and that all forms functional plasticity based on rapid and reversible
of perceptual learning can therefore be assumed to operate modifications of synaptic e‰cacy, although large-
within the framework of cortical adaptivity. scale amputations have been shown to involve
sprouting and outgrowth of a¤erent connections into
neighboring regions at cortical and subcortical levels
2.1 Introductory Remarks (Florence, Taub, and Kaas 1998; Jain et al. 2000).
Postontogenetic plasticity describes the capacity of Perceptual learning is often highly specific to stimulus
adult brains to adapt to internal or environmental parameters such as the location or orientation of a
changes. It is useful to distinguish between two dif- stimulus, with little generalization of what is learned
ferent forms of adult plasticity: to other locations or to other stimulus configurations
(see chapters 9, 11, 12, 14). Selectivity and locality of
1. Lesion-induced plasticity, which subsumes cor- this type implies that the underlying neural changes
tical reorganization after injury or lesion, induced are most probably occurring within early cortical
either centrally or at the periphery, refers to com- representations that contain well-ordered topo-
pensation for and repair of functions acquired before graphic maps to allow for this selectivity (see chapter
the injury or lesion. 1). In addition, a transfer of the newly acquired
2. Training- and learning-induced plasticity, often abilities is often considered an important marker of
called ‘‘use-dependent plasticity,’’ refers to plastic the processing level at which changes are most likely
changes that parallel the acquisition of perceptual occur: limited generalization is taken as evidence for
and motor skills. high locality of e¤ects in early representations. In
20 Hubert R. Dinse and Michael M. Merzenich
contrast, transfer of learned abilities is taken as evi- proximity (see chapters 14, 20). Because these changes
dence for the involvement of higher processing in input do not involve attention or processing for
levels often observed in task and strategy learning meaning, they induce a class of noncognitive adap-
(see chapters 13, 14). There is increasing evidence tations based largely on bottom-up processing.
that changes in early cortical areas might be more 2. By drawing attention to certain aspects of a
directly linked to perceptual learning than previously stimulus, thereby selecting it in comparison to others.
thought (Karni and Sagi 1991; Recanzone, Jenkins, The relevance of a stimulus can also change, depend-
et al. 1992; Schoups, Vogels, and Orban 1995; Crist ing on context, history, and behavioral task, thereby
et al. 1997; Fahle 1997, chapter 10). modifying how physically defined attributes are pro-
In fact, most of what we know today about adap- cessed. There is general agreement that modification
tation of the somatosensory system comes from the of early sensory processing by attention and stimulus
investigation of the somatosensory areas characterized relevance reflects top-down influences arising from
by extended and ordered neural representations of cognitive processes (see chapters 13, 14, 20).
the body surface (box 2.1). In contrast, less is known
3. By using reward or punishment to reinforce
about both the role of higher areas and the interac-
learning. Such influences usually accelerate adapta-
tion between sensory association areas for perceptual
tional processes and are assumed to be mediated
learning. In any case, the conjecture that perceptual
by specific brain regions modifying early sensory
learning a¤ects early areas provides an important
processing (see chapter 20).
conceptual link to somatosensory adaptational pro-
cesses (see chapters 9–14).
2.1.4 The Hebbian Metaphor
Box 2.1
Higher mammals contain complete and ordered topographic maps of the skin of the body surface giving rise to a
‘‘homunculus.’’ Adjacent locations on the skin are represented at adjacent locations in the cortex. Exceptions such as
face-hand border arise from the problem of mapping a three-dimensional object onto a two-dimensional surface. (a; d)
Examples of ‘‘homunculi’’ for monkey and rat. Monkeys are characterized by ‘‘multiple’’ representations. (b) Mirror-
like arranged body surface representations of areas 3b and 1 in an owl monkey. (c) Detailed map of the hand repre-
sentation often used for assessment of adaptational changes. In rats, most of our knowledge about the body
representation is from primary somatosensory cortex that is mapped along the dorsolateral aspect of the cortex. ( f )
Locations of hind paw, forepaw, and barrel system. (e) Video image of the dorsal surface of the brain, as indicated by
the thick dark line in panel f, with the locations of the hind and forepaws marked in white. (Panels a–c reprinted with
permission from Kandel et al. ( [1992] McGraw-Hill Companies; f reprinted with permission from Pellegrino et al.
[1986]).
22 Hubert R. Dinse and Michael M. Merzenich
must be extended beyond simultaneity in the sense training can transform a serial search to a parallel
of strict coincidence to cover all facets arising from search task (see chapters 13, 14, 20).
learning processes. Such a definition must include a On the other hand, there is little doubt about the
large number of pre- and postsynaptic patterns as significant contributing role of input statistics. Many
well as a broad time window of what neural systems studies have demonstrated that neural changes and
regard as ‘‘simultaneous.’’ parallel improvement of performance can be evoked
by a specific sensory input pattern without involving
2.1.5 Use-Dependent Plasticity as a Basis of attentional mechanisms, provided the statistics are
Perceptual and Motor Skills su‰ciently altered (see also section 2.4.2).
One of the striking features of use-dependent plas- 2.1.7 Top-Down Modulation of Plasticity
ticity is the correlation of cortical changes with per-
formance. The acquisition of skills has often been There are many brain centers that play a role in
used as an index for the buildup of implicit mem- modulating cortical responsiveness. The major source
ories. There are a number of crucial properties that of cholinergic inputs long implicated in learning
distinguish implicit from explicit memory. Implicit and memory comes from several groups of neurons
memories are acquired automatically and uncon- within the basal forebrain, which receives inputs
sciously. Many repetitions over a long time without from limbic and paralimbic structures. For example,
higher-level cognitive processes are su‰cient to in animal experiments, pairing of sensory stimulation
improve perceptual and motor skills. That the repe- with electrical stimulation of the nucleus basalis has
titions are noncognitive and many represents an been shown to result in rapid and selective reorga-
important aspect of use-dependent neural plasticity. nization (Rasmusson and Dykes 1988; Edeline et
It has therefore been speculated that use-dependent al. 1994; Bakin and Weinberger 1996; Bjordahl,
plasticity might be strongly related to, if not a sub- Dimyan, and Weinberger 1998; Kilgard and Merze-
strate for, implicit memory function. nich 1998). On the other hand, lesion of the chol-
inergic system has been shown to prevent plastic
2.1.6 Input Statistics versus Attention reorganization (Baskerville, Schweitzer, and Herron
1997; Sachdev et al. 1998). Consequently, chol-
As outlined above, attention plays an important role inergic inputs have been assumed to represent one
in learning and adaptational processes (Ahissar and example of a top-down system providing modu-
Hochstein 1993; Recanzone, Merzenich, et al. 1992; latory information of higher-order, presumably cog-
Goldstone 1998; Buchner et al. 1999). It has been nitive processes (cf. chapters 13, 14 and 20).
suggested that specific high-level attentional mecha-
nisms modify early sensory processing levels (Ahissar 2.1.8 Synopsis
and Hochstein 1993), although recent experiments
(Ito, Westheimer, and Gilbert 1998) indicate that at- The present chapter summarizes recent work on
tentional mechanisms themselves can be changed by somatosensory adaptations with special emphasis
practice (see also chapter 14). Similarly, researchers on behavioral and perceptual consequences of use-
(Sireoteanu and Rettenbach 1995) have shown that dependent plasticity as defined above. (For reviews
Adaptation of Inputs in the Somatosensory System 23
covering all facets of cortical plasticity, see Merzenich leads to a significant reduction of frequency dis-
et al. 1988; Kaas 1991; Garraghty and Kaas 1992; crimination threshold. When the cortical areas
Sameshima and Merzenich 1993; Donoghue 1995; representing the skin of the trained fingers were
Weinberger 1995; Cruishank and Weinberger 1996a; mapped, large-scale cortical reorganization became
Edeline 1996; Merzenich, Wright, et al. 1996; Dinse apparent, which included changes in receptive fields
et al. 1997; Kaas and Florence 1997; Sanes and and in the topography of cortical representational
Donoghue 1997; Buonomano and Merzenich 1998a; maps (Recanzone, Merzenich, et al. 1992). After
Nicolelis, Katz, and Krupa 1998.) training, sinusoidal stimulation of the trained skin
We first survey studies of training and use in a elicited larger-amplitude responses, peak responses
variety of animal models, then discuss recent studies earlier in the stimulus cycle, and temporally sharper
of somatosensory adaptations in humans by re- responses, than did stimulation applied to control
searchers using noninvasive imaging technologies. skin sites. Analysis of cycle histograms for area 3b
These imaging studies provide compelling evidence neuron responses revealed that the decreased variance
for the relevance of adaptational changes to everyday of each stimulus cycle could account for behaviorally
life. Next, we consider approaches that, by varying measured frequency discrimination improvements
input probability, explore how ‘‘driving factors’’ in- (Recanzone, Merzenich, and Schreiner 1992). These
duce adaptational changes. We review the increasing studies demonstrated for the first time a direct rela-
evidence for ‘‘maladaptive’’ aspects of neuroplasticity tion between cortical plasticity and improvement of
and touch on the role of ‘‘subcortical’’ plasticity. Fi- performance (see also figure 2.1), establishing a tight
nally, we critically examine what is changed during link between neurophysiological experiments and
adaptations, as this relates to the coding and decod- psychophysical tasks and making it possible to cor-
ing of sensory information during adaptations needed relate precisely defined aspects of use with plastic
to alter perceptual and motor performance. changes.
A related approach was taken by Xerri et al. (1999),
in which monkeys were trained to pick up food pel-
2.2 Role of Training, Di¤erential Use, and lets from wells of di¤erent sizes. Although all mon-
Behavior keys exhibited a gradual improvement in digital
dexterity, each monkey developed an individual re-
It is common wisdom that perceptual skills improve trieval strategy. In area 3b, the cortical magnification
with training (see Gibson 1953). Recent studies in of the di¤erentially engaged glabrous fingertip sur-
‘‘perceptual learning’’ have focused on problems and faces was nearly twofold larger than it was for con-
questions associated with skill acquisition. One of trol digits. Receptive fields of neurons representing
the most stimulating questions in cortical plasticity is the engaged digital surfaces were less than half as
how cortical changes are linked to changes in per- large as those representing the corresponding surfaces
formance, a question requiring simultaneous assess- of control digits. These results confirmed that behav-
ment of both neurophysiological and behavioral iorally important skin surfaces are represented in a
changes. much finer representational grain than normal.
For example, Recanzone, Jenkins, and coworkers Use-dependent plasticity has been investigated in
(1992) showed that tactile frequency discrimination a more natural context as well, where the link be-
training in adult owl monkeys over several months tween behavior and cortical reorganization is often
24 Hubert R. Dinse and Michael M. Merzenich
Figure 2.1
Adult owl monkeys were trained to detect di¤erences in the frequency of a tactile flutter vibration stimulus above a 20 Hz
standard. All stimuli were delivered to a constant skin site on a small segment of one finger. (a) Changes of psychophysical
performance in terms of threshold over successive sessions indicate progressive improvement with training that was highly
selective for the trained skin site. (b) Representative receptive fields defined in area 3b on the trained hand of the same
monkey shown in panel a. Filled circle denotes the area of skin trained in the behavioral task. The size of receptive fields
increased in the zone of representation of the trained digit as compared to adjacent digits or the control hand. (c) Cortical
representational map highlighting all penetrations that included some or all of the trained skin (dark stipple), the homolo-
gous skin on the adjacent digit (light stipple) or both skin surfaces (hatched). The inset shows the stimulated skin site (black
dot) and its equivalent on the adjacent finger (stippled dot). This analysis revealed that the representation of the stimulated
skin was larger than the control skin site. (d ) To study the temporal response characteristics, neurons were stimulated using
Adaptation of Inputs in the Somatosensory System 25
less quantifiable, although still intuitively obvious. ysis of receptive fields (RFs) and cortical representa-
In their study of lactating rats, Xerri, Stern, and tional maps of the hind paw revealed an increase of
Merzenich (1994) showed the implications of epi- RF size and a shrinkage of the cortical map within a
sodic di¤erential use following normal nursing be- few days. Both behavioral e¤ects and cortical changes
havior: the area SI representation of the ventral trunk were reversible within weeks ( Jürgens and Dinse
skin was significantly larger in lactating rats than in 1997a; Zepka, Jürgens, and Dinse 1996), demon-
matched postpartum nonlactating or virgin con- strating that even modest modification in behavior
trols. The greatest representational change was a can lead to rapid and large-scale cortical changes.
twofold increase of the cortical representation of the These findings extend the results of Fox and co-
nipple-bearing skin between the forelimbs and hind workers (1994), who studied adult rats under space
limbs. flight conditions after complete prevention from use
Housing rodents in an enriched environment or hind limb suspension. Both approaches resulted in
has been shown to inhibit spontaneous apoptosis, modified posture and gait, which returned to normal
prevent seizures, and produce general neuroprotec- after about two weeks. Behavioral adaptations were
tive e¤ects (Young et al. 1999). Furthermore, in the paralled by a reduction in the number of GABA-
hippocampus of even senescent mice, an enriched immunoreactive cells (D’Amelio et al. 1996).
environment induced neurogenesis (Kempermann, For the human motor system, similar fast ad-
Kuhn, and Gage 1997). The areal extent of the fore- aptational regulations have been reported. Using
paw cutaneous representation was significantly larger transcranial magnetic stimulation (TMS) mapping,
in rats housed in enriched environments promoting Licpert, Tegentho¤, and Malin (1995) have shown
di¤erential tactile experience (EE rats) for 71–113 that, in patients who had unilateral immobilization
days from weaning than in control rats housed under of the ankle joint without peripheral nerve lesions,
standard conditions (Coq and Xerri 1998). In addi- the area of motor cortex representing the tibial
tion, the receptive fields tended to progress in more anterior muscle was significantly smaller for the
orderly fashion across the digit glabrous skin of EE immobilized than for the una¤ected leg. The re-
rats than they did in control rats, corroborating the duction in area was correlated with the duration of
view that cortical cutaneous maps are maintained in immobilization, an e¤ect rapidly reversed by volun-
a permanent state of use-dependent fluctuation. tary muscle contractions.
In adult rats, a tendon of a hind limb was cut, The framework of modified use as a determinant
leading to slight changes in walking behavior. Anal- of cortical organization has been tested in an in-
sinusoidal tactile stimuli in the range of 20 to 26 Hz corresponding to the frequency range tested behaviorally. Upper
population cycle histogram (E-2) was constructed from summing neural responses of trained and stimulated digits (heavy
line) recorded at all area 3b locations, superimposed with histograms for stimulation of adjacent untrained digits (thin line),
in the monkey shown in panels a–c; bin width was 0.1 msec. Note entrained and faster response characteristics for the
trained digit, an e¤ect missing for the passively trained monkey shown in lower population cycle histogram (P-3). (e) Re-
gression analysis for the calculated threshold based on the overlap of the second cycle in the cycle histograms as a function of
behaviorally threshold, indicating that the training-induced temporal processing characteristics can explain behavioral per-
formance to a considerable degree. (Modified from Recanzone, Jenkins, et al. 1992; Recanzone, Merzenich, et al. 1992;
Recanzone, Merzenich, and Schreiner 1992. Reprinted with permission from the American Physiological Society.)
26 Hubert R. Dinse and Michael M. Merzenich
Adaptation of Inputs in the Somatosensory System 27
vestigation of age-related modifications of cortical Accordingly, in the case of neural degeneration, one
representational maps in old rats, known to show would expect comparable changes to occur in both
a number of age-related changes. The characteristic the forepaw and the hind paw representation. Indeed,
impairment of the sensorimotor state is most strik- analysis of RFs in the cortical forepaw representation
ingly expressed in a walking impairment of the hind of animals of high age revealed no alterations ( Jürgens
limbs (Schuurman et al. 1987; Ingram 1988; Stoll and Dinse 1997b). The results imply that age-related
et al. 1990). Using electrophysiological recordings, changes can be regionally very specific, thereby
researchers ( Jürgens and Dinse 1995; Spengler, arguing against an unspecific origin for such changes,
Godde, and Dinse 1995) demonstrated that behav- and that age-related neural changes and specific age-
ioral changes in old rats were paralleled by massive related behavioral alterations are linked.
reorganization of the somatosensory cortex (figure Interestingly, enriched environment has been
2.2). Age-related changes were characterized by an shown to prevent age-related decrease in synaptic
enlargement of receptive fields of the hind paw rep- density in the aged brain (Saito et al. 1994); indeed,
resentation, an increase of RF overlap and a deteri- when old rats were kept in an enriched environment
oration of the topography of the cortical maps. It has for several months, no comparable age-related alter-
been suggested (Dinse et al. 1995) that certain aspects ations of sensory (Churs et al. 1996) and motor hind
of these age-related behavioral changes reflect plastic paw representations (Reinke and Dinse 1999) were
changes resulting from prolonged disuse of the hind found. These results indicate that the beneficial out-
limbs rather than from age-related cortical degener- come of an enriched environment, namely, rein-
ation. Sensorimotor behavior of the forelimbs re- forced mobility and agility, occurs even in animals of
mains largely una¤ected even in animals of high age. high age.
Figure 2.2
Specific e¤ects of age on receptive fields of the hind paw recorded in somatosensory cortex of aged rats. (a, b) Representative
examples of behavioral changes of walking pattern derived from footprint analysis for young, control animal, left hind paw
and right forepaw and for old animal, left hind paw and right forepaw, respectively. Note selectivity of walking impairment
restricted to hind leg. Examples of receptive fields (RFs) recorded in the hind paw representation ( panels a and b, lower left)
and in the forepaw representation ( panels a and b, lower right) in young and old animal, respectively. Age-related changes
are limited to the behaviorally impaired extremity. To visualize the e¤ects of aging on the topography of the underlying
cortical maps, we reconstructed somatosensory maps using a computer-based interpolation algorithm based on a linear least
square approximation of sampling coordinates of penetration sites and corresponding receptive field centers. (c, d ) Recon-
structions of a cortical hind paw representation are shown for control and for old rat, respectively. Examples of cortical
topographies represented as a regular lattice within somatosensory cortex (left). Extrapolated cortical representation of a
schematic and standardized drawing of the hind paw (middle). Dashed lines indicate horizontal, and solid lines the vertical,
components of the lattice. One square of the lattice represents 1 mm 2 skin area. Diamonds indicate penetration sites; squares
give the interpolated RF centers. Dotted lines give the deviation between them. Backprojection of the regular lattice of the
cortical map onto the hind paw (right). Squares give the interpolated, and stars the measured, RF centers. One square of the
lattice represents the skin portions that is represented by 0.01 mm 2 cortical area. According to these reconstructions, maps of
the hind paw representation recorded in old animals, characterized by a selective impairment of the hindlegs show a dramatic
distortion of their representational maps and a loss of topographic order. (Modified from Spengler, Godde, and Dinse 1995;
Jürgens and Dinse 1997b. Panels c and d reprinted with permission from Spengler et al. ( [1995] Lippincott Williams &
Wilkins.)
28 Hubert R. Dinse and Michael M. Merzenich
Taken together, these studies suggest that small Furthermore, cortical reorganization of the finger
alteration in behavior due to special demands im- representation extending several millimeters was
posed in everyday life alters early cortical representa- observed in adults studied before and after surgical
tions rapidly and reversibly. The summarized studies separation of webbed fingers (syndactyly; Mogilner
imply that wearing a cast for some weeks due to a et al. 1993), a finding reminiscent of what had been
broken limb should be su‰cient to alter associated reported some years ago for artificial induction of
maps of somatosensory and motor cortex (for discus- syndactyly in monkeys (Clark et al. 1988).
sion of perceptual consequences, see section 2.3). Taken together, human studies confirm the close
From a more general point of view, the findings relation between intensified use and enlargement of
on age-related changes in the somatosensory system associated cortical representational maps thus sup-
extend the concept of use-dependent plasticity to porting the relevance of the concept of cortical plas-
high age. There is growing evidence from human ticity for everyday life.
studies that sensorimotor processing is more closely What are the functional implications of these
related to cognition than previously thought (Grady changes? Although, as discussed above (see section
and Craik 2000). Aging gives rise to an increasingly 2.1.5), the observed e¤ects are assumed to be the
strong association between sensory and cognitive substrate mediating the altered performance, there
functioning (Baltes and Lindenberger 1997). is controversy about the specificity of the neural
changes that accompany perceptual changes. Ac-
cording to one view, the adaptational changes are
2.3 Relevance of Adaptational Changes from
highly specific, allowing for improvement of the
Noninvasive Imaging Studies in Humans trained motor or perceptual skill only: neural changes
arising during training are assumed to have little e¤ect
The recent development of noninvasive imaging on information processing beyond that skill. Accord-
techniques has made it possible to study the impact
ing to an alternative view, neural changes result in a
of modified use and practice in humans. (For an
widespread modification of sensory processing over-
overview of the modern imaging techniques cur-
all; changes in perceptual and cognitive skills gener-
rently employed in human studies, see, chapter 7,
alize widely beyond the trained task. In other words,
this volume.) Imaging studies have provided over-
there is controversy over whether a specific im-
whelming evidence that extensive use and practice
provement is paralleled by other perceptual changes,
result in substantial changes of associated cortical
independent of the trained performance. There is
representations in blind Braille readers (Pascual-
evidence, for example, that many aspects of auditory
Leone and Torres 1993; Sterr et al. 1998a,b), in processing in blind subjects are superior to those in
players of string instruments (Elbert et al. 1995), in
normally sighted subjects (see Hollins 1989; see also
other musicians (Pantev et al. 1998), and in subjects Röder, Rösler, and Neville 1999, 2000). Pascual-
given long-term perceptual training in tactile dis- Leone and Torres (1993) reported increased sensory
crimination (Spengler et al. 1997). Of particular in- representation of the reading finger in blind Braille
terest are findings on cross-modal plasticity in blind readers, but no change in their spatial two-point
subjects (Sadato et al. 1996; Cohen et al. 1997, 1999; discrimination abilities, whereas Axelrod (1959)
Röder, Rösler, and Neville 1999; Röder et al. 1999). found evidence of improved discrimination perfor-
Adaptation of Inputs in the Somatosensory System 29
mance in such readers. More studies are needed to Spengler and Dinse 1994; Sil’kis and Rapoport 1995;
resolve this controversy and to clarify how far per- Gu and Fortier 1996; Kimura, Melis, and Asanuma
ceptual changes extend across di¤erent forms of 1996; Maldonado and Gerstein 1996a,b; Joublin et al.
specific adaptations and skills (see also section 2.8). 1996; Xing and Gerstein 1996; Dinse et al. 1997).
Although extremely beneficial in revealing sig- ICMS allows researchers to investigate the properties
natures of cortical plasticity, human studies of the of functional plasticity locally—independently of the
type discussed cannot determine the ‘‘driving factors’’ peripheral and subcortical pathways and indepen-
behind cortical reorganization. The exact nature of dently of the constraints provided by particularities of
inputs in a physical stimulus pattern is di‰cult, if not a sensory pathway and its preprocessing. In a typical
impossible, to assess. What is needed, therefore, are ICMS experiment, repetitive electrical pulse trains
complementary studies that investigate cortical reor- of very low currents (usually less than 10 mA) are
ganization induced by a systematic variation of input delivered via a microelectrode. Based on theoretical
pattern. calculations, ICMS of that intensity was assumed to
activate a cortical volume of only 50 microns in
diameter (Stoney, Thompson, and Asanuma 1968)
2.4 Role of Input Statistics
supporting the locality of the changes. Synchronized
discharges are generated, which are assumed to play a
Although many lines of evidence have shown that
crucial role in mediating plastic changes. The short
the somatosensory system adapts to input pattern
timescale and reversibility of ICMS e¤ects supported
of di¤erent probabilities, whether—without cog-
the hypothesis that modulations of synaptic e‰ciency
nitive processes—variation of input statistics alone in neuronal networks occur very rapidly without
su‰ces to reorganize cortical maps is a matter of long-
necessarily involving anatomical changes.
standing debate. Accordingly, studies that directly In the rat motor cortex, border shifts in move-
address the role of input probabilities for adaptational
ment representations exceeding 500 microns were
processes are needed to provide insight into principles
observed after a few hours of ICMS (Nudo, Jenkins,
and constraints governing adaptational processes.
and Merzenich 1990). Application of ICMS in the
hind paw representation of the adult rat somato-
2.4.1 Intracortical Microstimulation
sensory cortex caused an overall but selective expan-
sion of receptive field size (Recanzone, Merzenich,
A technique to evoke selective motor responses by
and Dinse 1992; Dinse, Recanzone, and Merzenich
applying current through microelectrodes inserted
1993; Spengler and Dinse 1994). Receptive fields
into defined regions of motor representations, intra- close to that of the stimulation site were enlarged, and
cortical microstimulation (ICMS) has more recently
comprised large skin territories always including the
been employed to study short-term and reversible
RF at the ICMS site, revealing a distance-dependent,
plastic changes in motor, somatosensory, auditory,
directed enlargement toward the ICMS receptive
and visual cortex as well as thalamic relay nuclei
field. Early ICMS-related reorganization could al-
of the somatosensory system (Nudo, Jenkins, and
ready be detected after 15 minutes of ICMS, and
Merzenich 1990; Dinse, Recanzone, and Merzenich
much greater e¤ects emerged after 2 to 3 hours.
1990, 1993; Recanzone, Merzenich, and Dinse 1992;
Changes were reversible within 6 to 8 hours after
30 Hubert R. Dinse and Michael M. Merzenich
termination of ICMS (Dinse, Recanzone, and Mer- In Diamond, Armstrong-Jones, and Ebner 1993,
zenich 1993; Spengler and Dinse 1994). sensory experience was altered by a few days of
Neural groups and assemblies are thought to be ‘‘whisker pairing’’: whiskers D2 and either D1 or D3
subject to modification during reorganizational pro- were left intact, whereas all other whiskers were
cesses. Using ICMS, researchers can directly address trimmed. Reorganizational changes were assessed in
the question of dynamic changes of neural assembly terms of response amplitude evoked by single neu-
membership (cf. 2.7). ICMS resulted in a significant rons recorded in the corresponding barrel field rep-
enhancement of correlated, synchronized neural resentations of somatosensory cortex. During the
activity that paralleled changes of cortical RFs and period of whisker pairing, the RFs of cells in barrel
cortical maps (Dinse, Recanzone, and Merzenich D2 changed in distinct ways: the response to the
1990, 1993). Similar results have been obtained for center RF, whisker D2, increased. The response to
plastic changes in auditory cortex (Maldonado and the paired surround RF whisker nearly doubled, and
Gerstein, 1996a,b). Where cortical neurons exhibited the response to all trimmed, unpaired surround RF
highly synchronous oscillatory firing patterns that whiskers decreased. After whisker pairing, the re-
were enhanced by ICMS, depending on the anato- sponse to the paired neighbor of D2 was more than
mical distance between the two neurons: ICMS twice as large as the response to the trimmed neigh-
changed the strength and the local number of such bor of D2. These findings indicate that a brief change
correlations. Overall, the results obtained with intra- in the pattern of sensory activity induced by pairing of
cortical microstimulation demonstrated the capacity tactile stimuli can alter the configuration of cortical
for cortical plasticity in the absence of peripheral RFs of adult animals.
stimulation. To test the hypothesis that consistently non-
coincident inputs may be actively segregated from
2.4.2 Coactivation Studies one another in their distributed cortical representa-
tions, monkeys were trained to respond to specific
To study the e¤ects of input statistics systematically, stimulus sequence events (Wang et al. 1995). Ani-
a number of protocols have been introduced in which mals received temporally coincident inputs across
the neural activity needed to drive plastic changes fingertips and fingerbases, but distal versus proximal
was generated by simultaneous, associative pairing digit segments were noncoincidentally stimulated.
(Diamond, Armstrong-Jones, and Ebner 1993; Wang Electrophysiological mapping experiments in area 3b
et al. 1995; Godde, Spengler, and Dinse 1996; showed that synchronously applied stimuli resulted
Godde et al. 2000). In a pioneering study by Frégnac in integration of inputs in the cortical maps, whereas
et al. (1988), functional persistent changes in re- stimuli applied asynchronously were segregated. In
sponse properties of single neurons of cat visual cor- contrast to those in normal control animals, cortical
tex were induced by a di¤erential pairing procedure, maps in trained animals were characterized by two
during which iontophoresis was used to artificially bandlike zones in which all neurons had multiple
increase the visual response for a given stimulus and digit RFs representing the stimulated skin surfaces.
decrease the response for a second stimulus. In con- These two zones were separated by a region con-
trast, the protocols described below used a pairing of taining normal single-digit RFs. This segregation
adequate (tactile) stimuli. was further augmented by a band of units responding
Adaptation of Inputs in the Somatosensory System 31
Figure 2.3
Typical reorganized cortical map of a hand in area 3b and receptive fields derived in an owl monkey engaged in a behavioral
training to discriminate tactile stimulus sequences delivered to the hand by two narrow bars. One bar stimulated a narrow
line of skin across the distal segments of the digits 2, 3, and 4; the second bar excited a narrow line of skin crossing the
proximal segments of the same three digits (thin lines in panel b). The training resulted in temporally coincident inputs
across fingertips and finger bases, whereas the distal versus proximal digit segments were noncoincidentally stimulated. (a)
The reorganized map shows that in contrast to normal maps, a significant portion of the map exhibited multiple receptive
fields that were specific to either the proximal (horizontal striping) or distal (vertical striping) portions of the digit. Inter-
estingly, both regions representing the trained skin surfaces became segregated by a band of single-digit receptive fields
(white) and a band of receptive fields located on the dorsum of the hand (black). (b) The corresponding receptive fields are
sorted according to the four classes observed: distal multiple-digit, proximal multiple-digit, dorsum, and single-digit recep-
tive fields. (Modified from Wang et al. 1995; Buonomano and Merzenich 1998a. Reprinted with permission from the
Annual Review of Neuroscience, vol. 21 ( 1998 by Annual Reviews www.AnnualReviews.org. Physiological Society.
Reprinted by permission from Nature ( 1995 Macmillan Magazines Ltd.)
to dorsal skin inputs, both features are normally not of adult rats were used for coactivation in order to
present in area 3b finger representations (figure 2.3). study e¤ects of simultaneous stimulation. The authors
Interestingly, maps derived in the ventroposterior reported reversible reorganization, namely, selective
portion of the thalamus (VPL) were not equivalently enlargement of the cortical territory and of the re-
reorganized suggesting that this particular type of ceptive fields representing the stimulated skin fields,
representational plasticity appears to be cortical in as well as emergence of a large representation that
origin. included a joint representation of both sites (figure
In Godde, Spengler, and Dinse 1996, non- or only 2.4). A control protocol of identical stimulus patterns
partially overlapping receptive fields on the hind paw applied to only a single skin site evoked no changes,
32 Hubert R. Dinse and Michael M. Merzenich
Figure 2.4
Scheme of the tactile coactivation protocol as utilized by Godde et al. (1996, 2000). (a) Two locations on the skin were
simultaneously stimulated for several hours with computer-controlled tactile stimulators. Example of the time course of
stimulation is shown. Average frequency is about 1 Hz. (b) Changes of receptive fields (RFs) as compared to control con-
dition (top left) after coactivation (bottom left) performed on d2 and d4 are characterized by a dramatic enlargement of RF
size. Note that, after coactivation, digit RFs cover always all of the skin surface of the coactivated digits d2 and d4. Cortical
maps derived before (top right) and after (bottom right) coactivation on d2 and d4. Black lines indicate blood vessels, pen-
etration sites are marked. Numbers indicate digits 1 to 5, p, pads; he, heels. Bars indicate locations where cells could not be
driven by low threshold cutaneous inputs. Cortical territories representing d2 or d4 are denoted by di¤erent grays. Dark gray
indicates zone of overlap between both representations, that is, a common representation of d2 and d4. After coactivation,
cortical territory representing the coactivated skin sites increased significantly; new skin representations containing the
stimulated skin fields emerged up to 400 microns beyond the control boundaries, whereas recording sites that were more
rostral maintained their unresponsiveness. Note also the emergence of a large common representation of both skin sites not
present under control conditions. (c) Tactile two-point discrimination thresholds of the index finger of the right hand as
measured in a two-alternative forced-choice discrimination experiment in thirty-five right-handed human subjects. Thresh-
Adaptation of Inputs in the Somatosensory System 33
indicating that coactivation was essential for induc- ments of the thumb and foot resulted in a reduction
tion. The selective and local changes within the cor- of the distance of the center of gravity of their respec-
tical map imply that early sensory cortical processing tive output maps in area MI as measured by TMS,
was a¤ected: only those areas undergoing a specific whereas asynchronous movements evoked no signifi-
alteration in input reorganized. cant changes, indicating that similar principles of coac-
To demonstrate the perceptual relevance of neural tivation hold for both the sensory and motor system.
changes induced by the coactivation protocol, Godde Recanzone, Merzenich, et al. (1992) recorded
et al. (2000) investigated spatial discrimination per- only modest increases in topographic complexity in
formance in human subjects given similar passive cortical hand representations for passively stimulated
costimulation of the tip of the index finger. Using hands and no e¤ects on RF size or overlap. On the
discrimination thresholds to mark reorganizational other hand, the coactivation studies discussed above
e¤ects due to variation of input statistics on human showed a clear e¤ect on both cortical and perceptual
perception, they found that two hours of coactiva- levels, even though attention was not involved.
tion su‰ced to drive a significant improvement of In the human discrimination experiments, subjects
the spatial discrimination performance. These results were instructed not to attend the stimulation; in-
demonstrated the potential role of pure input statis- deed, during the several hours of coactivation all
tics in inducing cortical plasticity without involving subjects continued their normal business work.
cognitive factors such as attention or reinforcement Conceivably, the engagement in normal day work
(figure 2.4; see also section 2.8). A combined assess- had not been possible without the simultaneous at-
ment of discrimination thresholds and recording of tentive engagement in other perceptual and motor
somatosensory evoked potentials in human subjects tasks. One explanation is that during the coactivation
revealed that the individual gain of discrimination protocol, which was, on average, applied at a rate of
performance was correlated with the amount of 1 Hz for several hours, selected skin regions were
cortical reorganization in primary somatosensory stimulated 10,000 times or more, a much higher
cortex as referred from the shifts of the location of frequency of stimulation than that given the
the N20 dipole (Pleger et al. 2001). monkeys during the passive discrimination training.
The coactivation protocol allows complete con- Conceivably, the intensity of the stimulation proto-
trol of timing and amount of stimulation. Using col might be the crucial factor responsible for its
asynchronous rather than synchronous stimulation e¤ectiveness. As stated in section 2.1.6, adaptational
resulted in reorganizations characterized by a large changes and parallel improvement of performance
separation of both stimulated skin sites (Zepka, can be evoked by changes in input statistics provided
Godde, and Dinse 2000). the statistics are su‰ciently altered.
Similarly, Liepert, Terborg, and Weiller (1999) Coactivation studies of this type are instrumental
reported that about an hour of synchronous move- in providing insight into the role of timing parameters
olds were measured five days before and immediately after coactivation (arrows) and on two subsequent days. In all subjects,
thresholds were reduced immediately after coactivation but returned to control values one day after coactivation, revealing a
time course of reversibility similar to that described for the electrophysiological reorganizational changes in rats. Control
experiments performed on the index finger of the opposite hand, which received no coactivation, showed no changes.
(Modified from Godde, Spengler, and Dinse 1996. Reprinted with permission. ( [1996] Lippincott Williams & Wilkins.)
34 Hubert R. Dinse and Michael M. Merzenich
that control reorganization in animals and humans: tients with chronic pain relative to that elicited by
they allow researchers to experimentally assess the painful back stimulation in healthy controls. Fur-
time window that defines simultaneity; to relate thermore, this enlargement showed a linear increase
neural changes as measured in animals with psycho- with chronicity of pain (Flor et al. 1997), suggesting
physical performance as assessed in humans; and that the cortical reorganization that accompanies
thereby to explore the perceptual relevance of sim- chronic pain may serve an important function in its
ple paradigms that successfully induce adaptational persistence.
changes at the neuron level. Repetitive strain injuries including occupationally
induced focal dystonia have a high prevalence in
workers who perform heavy schedules of rapid al-
2.5 Therapeutic Consequences of ternating movements or repetitive, sustained, coor-
Somatosensory Adaptations dinated movements. It has been hypothesized that
use-dependent plastic changes such as those reviewed
The final outcome of reorganizational process need above may cause repetitive strain injuries charac-
not be beneficial. There is increasing evidence that terized by sensory dysfunction and impairment of
abnormal perceptual experiences such as the phantom motor control (Byl, Merzenich, and Jenkins 1996;
limb sensation arise from reorganizational changes Byl et al. 1997). This view is supported by studies
induced by the amputation of the limb. A strong re- showing that monkeys trained in repetitive hand
lationship has been reported, but not for nonpainful closing and opening developed typical movement
phantom phenomena experienced after arm ampu- control disorders indicated by a 50% drop in motor
tation, implying that these changes are maladaptive performance. Electrophysiological mapping within
rather than adaptive (Flor et al. 1995). By contrast, the primary somatosensory cortex revealed a dedif-
neuromagnetic source imaging revealed minimal ferentiation of cortical representations of the skin of
reorganization of primary somatosensory cortex in the trained hand, manifested by receptive fields 10 to
congenital amputees and in traumatic amputees not 20 times larger than normal, as well as by a break-
su¤ering from phantom limb pain (Flor et al. 1998). down of the receptive field topography (Byl, Mer-
These data indicate that phantom limb pain is related zenich, and Jenkins 1996). Thus repetitive, highly
to, and may be a consequence of, plastic changes in stereotypic movements can actively degrade cortical
primary somatosensory cortex. Interestingly, recent representations of sensory information guiding fine
studies reported a decrease of phantom limb pain motor hand movements. Using MEG, Elbert et al.
associated with prosthesis-induced use of the stump (1998) found a smaller distance between the repre-
(Weiss et al. 1999) and with less related reorganiza- sentations of the digits in somatosensory cortex for
tion in the motor cortex (Lotze et al. 1999). The the a¤ected hand in musicians su¤ering from focal
precise topographic mapping of the phantom limb hand dystonia than for the hands of nonmusician
onto the face area was explained in terms of the control subjects, indicating that central reorganiza-
topography of the border of the face-hand maps tion develops as a consequence of repetitive strain
(Ramachandran, Stewart, and Rogers-Ramachandran injuries in humans as well (see also chapter 7).
1992; Halligan et al. 1993; Aglioti et al. 1997). The maladaptive consequences of cortical plasticity
The power of the early evoked magnetic field have become increasingly acknowledged as a major
elicited by painful stimulation was elevated in pa- factor in various forms of dysfunctions. For example,
Adaptation of Inputs in the Somatosensory System 35
a negative outcome of neuroplasticity may play a at a subcortical level may depend on feedback con-
major role in some forms of age-related changes. nections from the cortex (Ergenzinger et al. 1998;
Jürgens and Dinse (1997b) found that walking Krupa, Ghazanfar, and Nicolelis 1999). Further-
impairments that develop in rats of high age as a more, the usually extensive cortical reorganizations
secondary response to muscle atrophy and to other may in part depend on activation of the widespread
factors promoting limited agility were the result of horizontally connected network that is lacking in
maladaptive cortical reorganizations. thalamic nuclei.
Studies of subcortical and brain stem structures have Earlier work on cortical reorganization concentrated
clearly shown that reorganizational changes occur largely on the analysis of receptive fields and on the
along the entire sensory pathway ( Wilson and Snow areal extent of representational maps. Both variables
1987; Garraghty and Kaas 1991; Nicolelis et al. still figure widely in current research, and abundant
1993; Pettit and Schwark 1993; Florence and Kaas data have allowed researchers to make valuable com-
1995; Faggin, Nguyen, and Nicolelis 1997; Jones parisons between di¤erent species and modalities
and Pons 1998; Nicolelis, Lin, and Chapin 1997; (box 2.2). In addition, new imaging techniques, such
Hubscher and Johnson 1999; Xu and Wall 1997, as fMRI, have allowed them to study adaptation-
1999; Woods et al. 2000). On the other hand, cor- al changes in humans more precisely, to describe
tical changes have been reported in the absence of changes in neural representations in terms of the
parallel subcortical changes. Accordingly, there are activation size of cortical maps, and thereby to link
still many open questions about the role of subcortical human and animal studies. Temporal aspects of
substrates. For example, Wang and coworkers (1995) coding have become increasingly prominent; conse-
did not find topographic changes in the thalamic relay quently, synchronicity and correlated activity have
nucleus comparable to those in area 3b after co- been intensively studied (see 2.4.1, figure 2.5). There
activation, as described in section 2.4.2. Similarly, is now substantial evidence that cooperativity among
Tinazzi et al. (1997) could not find parallel sub- many neurons is subject to modification during
cortical changes after transient dea¤erentation in plastic reorganization, implying that changes in
humans. Dinse et al. (1997) found that applying temporal coding are important or even crucial to
microstimulation in the thalamus had little e¤ect on use-dependent plasticity. This evidence supports the
RF size when compared to applying such micro- hypothesis that neural assembly membership is
stimulation directly in cortex. organized along primarily dynamic lines (for theo-
Of course, the central question behind these studies retical work, see Edelman and Finkel 1984; Braiten-
is exactly what role cortices play in the overall pro- berg 1986; von der Malsburg 1987; Aertsen et al.
cesses of reorganization (Darian-Smith and Gilbert 1989; Abeles 1991; Shenoy et al. 1993; for experi-
1995; Fox 1994; Florence, Taub, and Kaas 1998; mental work, see Dinse, Recanzone, and Merzenich
Kaas and Ebner 1998; Kaas 1999; Kaas, Florence, 1990, 1993; Ahissar et al. 1992, 1998; Nicolelis et al.
and Jain 1999). Much of the plasticity encountered 1993; Nicolelis, Lin, and Chapin 1997; Maldonado
36 Hubert R. Dinse and Michael M. Merzenich
Box 2.2
Two di¤erent ways to assess cortical maps, activity distributions, and receptive fields. (a) Action potentials from single
neurons or from small clusters of neurons are measured by inserting microelectrodes into the middle layers of the
cortex. The receptive field on the skin surface is defined as that region where stimulation evokes action potentials.
This procedure maps activity recorded in the cortex into the stimulus space that allows an easy and systematic way of
parametric analysis. When moving the electrode to an adjacent location in the cortex, a systematic shift in the cor-
responding receptive field location will be encountered. A complete topographic map can be obtained when a large
number of electrode penetrations are combined such that the penetration coordinates are related to the corresponding
receptive field coordinates. (b) Cortical activity distributions are measured. In contrast to panel a, a fixed stimulus,
ideally a small, ‘‘pointlike’’ stimulus, is applied, and the entire activity in the cortex evoked by that stimulus is mea-
sured. This type of activity distribution if often referred to as ‘‘point spread function’’ (PSF). Technologies used for
this kind of analysis are optical and functional magnetic resonance imaging. It should be noted, however, that the PSF
can be obtained using microelectrodes, in which case the activity of single or multiple neurons evoked by the
‘‘pointlike’’ stimulus is recorded and its spatial distribution is derived from a systematic mapping of that response at
di¤erent locations.
Adaptation of Inputs in the Somatosensory System 37
speech and speech stimuli, suggesting that the re- substructures. Accordingly, subtle but highly task-
organizational changes are specifically sensitive for specific and more complex changes in RF organization
temporal parameters of the input. (Tallal, Miller, and that are easy to miss experimentally may contribute
Fitch 1993; Merzenich, Jenkins, et al. 1996; Tallal et to an improvement of perceptual performance. On
al. 1996). the other hand, several potential mechanisms can
Thus adaptational processes alter both spatial and result in RF enlargement: downregulation of
temporal aspects of sensory processing, which has inhibitory surrounds; strengthening of excitatory,
implications for all forms of perceptual learning, and previously subthreshold RF regions; enhanced,
specifically for tasks involving spatiotemporal prox- synchronized discharge resulting in lowering of
imity such as interval or saltation learning (Geldard thresholds; or a mixture of all these putative mecha-
and Sherrick 1972; Cholewiak 1976; Kilgard and nisms. In any event, monitoring RF size without
Merzenich 1995). monitoring the mechanisms underlying changes in
RFs will yield only limited insight into the meaning
of the changes.
2.8 Coding of Adaptational Changes For the tactile coactivation protocol described
above, Godde and coworkers (Godde, Spengler, and
Changes in receptive fields, cortical maps, correlated Dinse 1996; Godde et al. 2000) observed an RF
activity, and temporal sequence processing (see figure increase in rats and a lowering of the discrimina-
2.7) clearly suggest that use and training alters cortical tion threshold in human subjects. Assuming that the
processing. Yet, in most cases, the causal link between coactivation protocol results in comparable changes
neural changes and changes in performance remains in both man and rat, the enhancement of the dis-
to be clarified. Recanzone, Merzenich, and co- crimination performance appears at first sight to
workers (1992) pointed out that there was no cor- contradict the reported receptive field enlargement.
relation between increased RF size and performance. A discrepancy between perceptual thresholds and
In contrast, the enhancement of temporally coherent single-neuron properties is not a new finding, how-
responses correlated strongly with the discrim- ever. For example, Westheimer (1979) found that
inative performances (Recanzone, Merzenich, and hyperacuity could not be explained in terms of the
Schreiner 1992). receptive field sizes of single cells. Coactivation-
Indeed, many studies have reported an increase in induced plasticity included enlargement of receptive
receptive field size in response to a broad spectrum fields, accompanied by increased receptive field
of di¤erent forms of induction of adaptational overlap and enlargement of the representational
changes. This frequently observed phenomenon maps, which, in turn, reflected an increase in the
makes it di‰cult to attribute a specific coding aspect total number of neurons activated by the stimulation
to RF size. It could be that strengthening of synapses and thus of processing resources. Temporal aspects of
during a Hebbian learning process inevitably leads neuron responses were also changed in terms of re-
to RF enlargement with little bearing on behavior. sponse duration and paired pulse behavior. It appears
From a methodological point of view, assessment of reasonable that all these changes, in concert, enable
RF size is usually done by hand-plotting techniques cortical networks to perform a faster and more elab-
that provide scant information about the internal orate decoding and processing of information.
organization of receptive fields and their spatial
40 Hubert R. Dinse and Michael M. Merzenich
Figure 2.7
Scheme of adaptational processes and interaction with information processing. The main driving forces of adaptations are
input statistics and context, on the one hand, and attention and meaning, on the other; by acting bottom-up and top-down,
respectively, these modify information processing. Among the parameters subject to modification by adaptational processes
are the typical descriptors of cortical processing such as receptive fields, tuning properties, and various aspects of temporal
processing and cooperativity among many neurons. In addition to training- or use-specific alteration of such parameters,
processing resources can be adjusted according to the requirements of the behavioral status.
In the psychophysical experiments performed by hart 1986; Baldi and Heiligenberg 1988; Eurich and
Godde and coworkers (2000), spatial discrimination, Schwegler 1997; Eurich et al. 1997) has been used to
but not localization abilities were tested. Evidence explain high-resolution performance by a population
for a trade-o¤ between localization and discrimina- of neurons with broad tuning characteristics: given
tion can be inferred from Sterr et al. 1998a, which su‰cient overlap between tuning curves, each de-
reported that stimuli were more often mislocalized sired resolution can be achieved. If we assume that it
on the reading fingers of three-finger Braille readers is not the property of a single cell that determines
than on control fingers (see chapter 7). This finding behavior, the coarse coding principle can be consid-
suggests that spatial discrimination performance ered a variant of the more general population coding
might benefit from enlarged receptive fields at the approach. Neural population analysis implies that
expense of localization performance. Thus further large ensembles of neurons contribute to the cortical
understanding of what is coded requires not only representation of sensory or motor parameters. Early
detailed analysis of neural changes, but also a broad formulations of this idea postulated that complex
battery of psychophysical and behavioral tests to find stimuli are represented by the simultaneous activa-
out what really is changed, what is improved, and tion of elementary feature detectors (see Erickson
what may become impaired. 1974). In primary motor cortex, ensembles of neu-
From a theoretical point of view, the ‘‘coarse rons broadly tuned to the direction of movement
coding’’ principle (Hinton, McClelland, and Rumel- have been shown to accurately represent the current
Adaptation of Inputs in the Somatosensory System 41
2.9 Conclusions
with artificial depolarization. Adult sensory cortices tatory RF or to change their subfield composition
(visual, somatosensory, and auditory) are capable within the RF of simple cells. By carefully mapping
of considerable plasticity (Kaas 1991; Eysel 1992; the RF with averaged post-stimulus-time histograms
Weinberger 1995; Cruikshank and Weinberger (PSTHs) in response to optimally oriented on-o¤
1996b; Gilbert 1998; see also chapter 1). Cellular stimuli, we were able to determine the exact size
correlates of long-term use-dependent changes in of the RF and to identify an adjacent completely
attentive learning have been shown in the auditory unresponsive region on one side of the RF as test
and somatosensory cortices of awake, behaving ani- region by repetitive stimulation and statistical evalu-
mals using natural sensory stimuli given over days to ation; another region, located just outside the RF or
weeks (Recanzone et al. 1992, 1993; Bakin and in the minimal discharge field at the opposite border
Weinberger 1990; see also chapter 2). To induce of the RF, was selected as control region so as to de-
changes in cellular responses in vivo on a faster tect even minimal unspecific changes (figure 3.1A).
timescale, natural sensory stimuli have been paired The experimental conditioning stimulation, which
with electrical or pharmacological stimulation of we called ‘‘visual costimulation,’’ consisted of repet-
cells in the visual, auditory, and somatosensory cor- itive 1/sec ‘‘on-o¤ ’’ stimuli covering parts of the
tices of anesthetized animals (Frégnac et al. 1988, excitatory RF (including its highly responsive cen-
1992; Cahusac 1995; Cruikshank and Weinberger ter) and the unresponsive test region just outside the
1996a; Frégnac and Shulz 1999). These experiments RF (figure 3.1B). In about one-third of all cells, a
have provided in vivo evidence that the modifiabil- specific widening of the RF into the unresponsive
ity of mature visual cortical connections follows the region was observed when the previously unrespon-
Hebbian learning rule, which requires both pre- sive region was tested 15 to 180 min after 15 min
and postsynaptic activity within a defined time win- of costimulation (figure 3.1C). In general, this e¤ect
dow (Hebb 1949). However, none of the above was observed after 10 to 75 min of costimulation
approaches answered the question as to whether and persisted from 20 min to more than 3 hours. In
LTP-like activity-dependent changes—as seen in none of these cases did we see a change of the re-
adult visual cortical cells in vitro (Artola and Singer sponse in the nonstimulated control region. In addi-
1987; Kimura et al. 1989; Hirsch and Gilbert 1993; tion to the associative plasticity seen at the border of
Kirkwood and Bear 1994)—can be obtained with receptive fields, potentiation of responses inside the
purely natural sensory stimulation in vivo. classical RF was often observed. For example, after
We approached this issue by recording from single costimulation, which elicited a strong ‘‘o¤ ’’ re-
cells in the visual cortex of the anesthetized adult cat sponse, an initially small ‘‘o¤ ’’ response in the central
and showed that long-lasting changes in receptive region of an excitatory RF increased and changed
field size and substructure can indeed be obtained the center response from a predominantly ‘‘on’’ to an
within minutes (Eysel, Eyding, and Schweigart ‘‘on-o¤ ’’ response after 60 min of repetitive stimu-
1998). To elicit ‘‘associative’’ synaptic changes, visual lation. We observed no di¤erences for any of these
costimulation was applied to the central parts of an e¤ects between areas 17 and 18 or between simple
excitatory RF and its immediate unresponsive sur- and complex cells. In addition, the findings were
round. Our aim was to stimulate visual cortical cells independent of RF size and eccentricity within 10
so that they would specifically ‘‘learn’’ to respond to around the area centralis.
an originally subthreshold region outside their exci-
Plasticity of Receptive Fields of the Adult Visual System 45
Figure 3.1
Receptive field expansion as consequence of associative costimulation of the classical receptive field with an unresponsive
adjacent region. (A) Schematic drawing of the classical receptive field (RF; oblique hatching), the unresponsive test region
(gray) on one side and the unresponsive control region (vertical hatching) on the other. The white bars represent the light
stimuli used for stimulation. The post-stimulus-time histograms (PSTHs) to the right show the missing response in the test
region and the strong ‘‘on’’ and weaker ‘‘o¤ ’’ response in the center of the classical RF prior to the experiment. Like the test
region, the control region shows no response. (B) The associative visual costimulation is applied for 15 minutes (900 ‘‘on-
o¤ ’’ stimuli). The stimulus covers a large part of the classical RF and the test region. The ‘‘on-o¤ ’’ response during
costimulation resembles the response from the RF center. (C) The expansion of the RF toward the costimulated test region
is proven by the clear ‘‘on’’ responses in the PSTHs obtained from that region 15 minutes and 3 hours after costimulation.
46 Ulf T. Eysel
intracortical circuits in gating of visual cortical plas- visual stimuli rapidly induced long-lasting RF mod-
ticity at the network level (Kirkwood and Bear ifications could provide the basis for modeling the
1994). Long-term potentiation was successfully in- cellular mechanisms that underlie this fast visual
duced in cat visual cortex slices by stimulation of learning, which is most probably rooted in early
long-ranging horizontal fibers when the elicited visual processing.
responses were predominantly excitatory, but not Although visual RFs are characterized by supra-
when the composite responses contained significant threshold inputs, subthreshold synapses within the
inhibitory potentials (Hirsch and Gilbert 1993). RF and close to the borders may also be activated.
In our first attempt (Eysel, Eyding, and Schweigart Associative coactivation of sub- and suprathreshold
1998) to show RF plasticity in anesthetized cat visual synapses—presumably converging on common so-
cortex in vivo by applying repetitive natural visual matic or dendritic sites—appears to lead to specific
stimulation, we exclusively investigated cells in layers synaptic strengthening and to locally changed RF
II/III shown to be capable of LTP in in vitro experi- substructure and enlarged RFs. Thus mature visual
ments (Artola and Singer 1987; Kimura et al. 1989; cortical cells can ‘‘learn’’ to respond to previously
Hirsch and Gilbert 1993; Kirkwood and Bear 1994) subthreshold inputs during synchronous coactivation
and used simple, unspecific large field conditioning in an anesthetized state in vivo, a correlate of per-
stimuli and optimally oriented test stimuli to look for ceptual learning at the level of the single cell or
significant results with a minimal number of experi- cortical network.
mental variables. To determine the specificity of the
e¤ect and whether it can be observed in cortical
layers other than II/III, further experiments will be 3.3 E¤ect of Dea¤erentation on Cortical
Receptive Fields and Retinotopy
needed, using di¤erent orientations and variable
temporal coherence of conditioning stimuli in dif-
Fast changes in visual receptive field properties also
ferent parts of the RF and recording from other
follow acute disuse associated with retinal lesions
cortical layers.
(Chino et al. 1992; Gilbert and Wiesel 1992) and are
Pathways that might be involved in the e¤ects
elicited when selective surround stimulation is ap-
observed here in layer II/III cells include collaterals
plied to artificial retinal scotomas (Pettet and Gilbert
of geniculocortical a¤erents (Gilbert and Wiesel
1992; Das and Gilbert 1995b; DeAngelis et al. 1995).
1979; Freund, Martin, and Whitteridge 1985) acting
Retinal lesions lead to an interruption of visual
via interneurons of the same cortical column; long-
signal flow from the destroyed region. Cells in the
range excitatory axons of cells in layers II/III/IV
(Gilbert and Wiesel 1979; Hirsch and Gilbert 1993; thalamus (lateral geniculate nucleus) and the primary
visual cortex are locally deprived of their normal
Kisvárday and Eysel 1992; see also chapter 1); and
input. Because of the predominantly binocular in-
interactions between areas 17 and 18 and higher
nervation of visual cortical cells, only homonymous,
visual areas.
binocular lesions can cause dea¤erentation of the
In psychophysical experiments with adult humans,
cortical cells. Circumscribed retinal lesions introduce
training has been shown to significantly improve per-
a characteristic imbalance in the target regions: cells
formance in visual perceptual tasks (Fiorentini and
with normal a¤erent input are situated in the im-
Berardi 1980; Fahle, Edelman, and Poggio 1995; see
mediate neighborhood of visually silenced cells.
also chapters 9–12). Our finding that repetitive
48 Ulf T. Eysel
This situation leads to changes in cortical cell center of an artificial retinal scotoma (a masked re-
RF size and topography when the active cells gain gion of the retina surrounded by moving pattern
influence on their inactive neighbors through the stimuli), significant increases of RF size were mea-
long-range intracortical horizontal fiber system sured (Pettet and Gilbert 1992; Das and Gilbert
(Darian-Smith and Gilbert 1995; Das and Gilbert 1995b). This is equivalent to a filling in of the masked
1995a). This possible basis for the perceptual filling region with information from the surround. Al-
in of ‘‘artificial’’ and ‘‘real’’ retinal scotomas will be though DeAngelis et al. (1995) questioned the RF
addressed in greater detail below. Such reorganiza- size increases reported in these studies, arguing that
tion after a retinal lesion occurs at the cortical level there was primarily a general increase in responsive-
with redundant information from the border of the ness and that normalization of the activity resulted
retinal lesions, that is, the response to a stimulus at in unchanged receptive field dimensions, further
the border of the lesion is represented in its correct experiments showed a genuine RF expansion, with
cortical location as well as laterally displaced in the cells responding to parts of the visual field, where
originally dea¤erented region (reorganization with previously they could not be excited (Gilbert et al.
redundant information; figure 3.3); it has been de- 1996).
scribed in the cat and monkey visual system (Kaas et Similar e¤ects were seen when unilateral or bilat-
al. 1990; Heinen and Skavenski 1991). eral retinal lesions were applied. Within minutes after
homonymous binocular retinal lesions, the RFs in
3.3.1 E¤ect of Artificial Scotomas and Acute primary visual cortex of monkeys and cats expanded
Retinal Lesions on Receptive Fields at the border of the cortical representation of the
lesions (Gilbert and Wiesel 1992). RFs reached, on
Psychophysical observations yielded an immediate average, five times their original area (growing from
filling-in phenomenon for retinal scotomas (Gerrits 0.07 to 0.37 degree 2 ); these early changes were also
and Timmermann 1969) and a delayed filling-in accompanied by shifts in retinotopy and map reor-
phenomenon for stabilized retinal images ( Yarbus ganization when retinal lesions were smaller than 5
1957; Millodot 1965; Gerrits, de Haan, and Vendrick of visual angle (Chino 1995, 1997).
1966) that can be regarded as ‘‘artificial retinal scoto-
mas’’ (Gerrits and Vendrick 1970). A stabilized image 3.3.2 Lesion-Induced Changes in Excitability in
follows all eye movements, hence the contours of its the Surround of the Scotoma
borders are not enhanced by the continuous motion.
Such an image vanishes with time, it is filled in from The retinotopic projections of homonymous retinal
the surround with redundant information similar to lesions ( panel A of figure 3.3) cause a cortical region
the (instantaneous) filling in of the blind spot or a devoid of retinal input from the homonymous visual
retinal scotoma. field area from both eyes (dea¤erented visual cortex;
In recent years, artificial scotomas were intro- panel B), which is associated with a blind region in
duced into experimental neurophysiology (Pettet the visual field (scotoma; panel C). To measure the
and Gilbert 1992). Short-term RF size changes were changes of cortical excitability introduced by the
found in the cells within an artificial scotoma. When retinal lesions, we recorded spontaneous activity and
single-cell receptive fields were recorded from the visually evoked activity inside and outside the deaf-
Plasticity of Receptive Fields of the Adult Visual System 49
Figure 3.3
Methods and basic topographical reorganization a month after bilateral central retinal lesions. (A) Fundus photographs of
homonymous 10 diameter retinal lesions in the right (R) and left (L) eye. (B) Schematic drawing of the lesioned eyes and
one hemisphere of the visual cortex. The bilateral, homonymous lesions in the eyes (black rectangles) lead to a region of
dea¤erentation in the visual cortex (gray area). The electrode array covers the dea¤erented region and a part of cortex an-
terior to it. The individual electrodes are spaced by 1 mm and are numbered from 1 to 7 from posterior to anterior. Fields
are mapped with an array of recording electrodes (1–7). (C) The receptive fields (RFs) of cells recorded in the visual cortex
are drawn in the contralateral visual field. The four anterior electrodes (4–7) are located outside the dea¤erented region and
have receptive fields migrating upward along the vertical meridian (VM) up to the border of the retinal scotoma. The three
posterior electrodes (1–3) are located within the dea¤erented cortex. RFs 2 and 3, obtained at these cortical locations, are
displaced in the visual field and are found at the border of the scotoma.
50 Ulf T. Eysel
weeks. Accordingly, the RFs of these reconnected The glutamic acid decarboxylase (GAD) immuno-
cells change their retinotopy and shift to a new reactivity, though downregulated inside the scotoma
position located at the border of the retinal lesion such that the number of positive profiles in the neu-
(Eysel, Gonzalez-Aguilar, and Mayer 1980, 1981). ropil was extremely reduced, remained essentially
This reorganization leads to a partial filling in of the unchanged in the cell somata (Rosien et al. 1995).
scotoma on the subcortical level (Eysel, Gonzalez- The immunohistochemistry of the excitatory neuro-
Aguilar, and Mayer 1981) and is associated with a transmitter glutamate revealed glutamate-positive
lateral spread of excitation in the dLGN by up to cells in cortical layers II to VI of area 17. The retinal
300 mm (Eysel 1982). With the same type of lesion lesions caused a clear reduction (by 18–26%) in the
(combined with enucleation of the other eye), Kaas number of glutamate-immunoreactive cells in the
et al. (1990) observed a much more extended reor- supra- and infragranular layers of the cortical area
ganization in cat visual cortex. Apart from the longer representing the lesioned retina, compared to nor-
distances, the reorganization was exactly of the same mal cortex. Furthermore, the cortex just outside
kind as observed in the dLGN: originally silenced the dea¤erented region displayed a sharp increase (of
cells regained visual input that originated from the 50–100%) in glutamate immunoreactivity through-
border of the retinal lesion, a retinal region that had out layers II to VI of area 17, with the largest peak,
not excited these cells before. The same kind of having a width of 600–800 mm, noted in layer VI
plasticity was found in monkey visual cortex (Hei- (Arckens et al. 2000).
nen and Skavensky 1991). In cat and monkey, the Both the changes in GAD and glutamate immu-
reorganization followed a characteristic time course, noreactivity diminished with time. When the central
filling in cortical regions that were depleted of reti- and the peripheral portions of area 17 in cats with
nal inputs by homonymous retinal lesions up to postlesion survival times of longer than twelve weeks
10 mm across, equivalent to an intracortical lateral were compared, no significant di¤erences in the
spread of excitation of 5 mm (Gilbert and Wiesel number of glutamate immunoreactivity-positive cells
1992). At the same time, a significant enlargement of and in the GAD immunoreactivity of the neuropil
RFs was observed that occurred quite early after were observed, showing that the immunohisto-
lesioning (for reviews, see Chino 1995, 1997). chemical reactions return back to normal once the
functional reorganization is completed.
3.3.4 Cortical Immunohistochemistry after Chronic
Retinal Lesions
3.4 Cortical Lesions and Perilesional Cortical
When the immunohistochemistry (IHC) of the Reorganization
GABAergic system (glutamic acid decarboxylase
Retinal lesions that switch o¤ the a¤erent input to
IHC; Rosier et al. 1995) and of the glutamatergic
the cortex are quite di¤erent from cortical lesions
system (glutamate IHC; Arckens at al. 2000) was
that destroy target cells for the retinal projection to
investigated two weeks after homonymous lesions in
the cortex; they also pose quite di¤erent problems
the retina (central 10 ), researchers discovered, inside
for reorganizing, and restoring function to, the visual
and in the surround of the dea¤erented cortical re-
system. As described above, in the case of a retinal
gion, a characteristric pattern of transmitter system
lesion, the system must reinnervate cortical cells that
immunoreactivity induced by the retinal lesions.
52 Ulf T. Eysel
have lost their sensory input, using redundant infor- Schmidt-Kastner 1991) or by ibotenic acid injections
mation from the border of the scotoma. In the case (Schweigart and Eysel 1998). The lesions were
of a cortical lesion, the retina and a¤erent pathways placed in one hemisphere of the cat visual cortex at a
remain completely intact, but the cortical target cells point some 3 mm lateral to the midline (area 17,
used to represent a certain retinal topography are sometimes extending into area 18). The typical le-
lost. The problem for reorganizing the system is sion completely destroyed the supragranular layers
completely di¤erent: surviving cells in the surround and reached down into the infragranular layers (see
of the lesion must take over the topographical rep- figure 3.5A). The histological changes observed in the
resentation of their destroyed neighbor cells; ac- area surrounding these lesions are described in greater
cordingly, the functional outcome is fundamentally detail elsewhere (Eysel 1997; Schmidt-Kastner et al.
di¤erent because information is again made accessi- 1993; Mittmann et al. 1994; Schroeter et al. 1995).
ble for cortical processing from retinal regions no When we recorded single cells in anesthetized cats
longer represented (reorganizing with nonredundant 1, 2, 7, and 30 days after cortical lesioning, we
information). A chronic reorganization of this type observed typical changes 1–2 days after the lesion.
has recently been described in cat striate cortex The spontaneous and visually driven activity of the
(Eysel and Schweigart 1999). Interestingly, the two cells was dependent on the distance from the border
completely di¤erent types of lesions seem to trigger of the cortical lesions. Activity was suppressed close
common early e¤ects leading to receptive field plas- to the border of the lesion (up to 0.5 mm), increased
ticity that appears useful for a certain recovery of at distances between 1 and 1.5 mm, and approached
function (Eysel et al. 1999). normal values farther away from the lesion (figure
The lesion-induced changes can be roughly sub- 3.5B). Seven to 30 days after the lesion, spontaneous
divided into acute, subacute, and chronic e¤ects. The excitability returned to normal values both for the
pathology predominant in the acute phase (first post- depressed cells close to the lesion and for the hyper-
lesion day) is followed by events of neuronal plasticity excitable cells in the adjacent region.
in the subacute phase (two days to one week after a In the acute and subacute phases, in vivo mapping
lesion) and the chronic phase (weeks to months after of single cells using extracellular microelectrodes
the lesion; Eysel 1997) that can be useful for mini- in cats with focal heat lesions of the visual cortex
mizing the functional loss caused by the lesions. yielded activity in the surround of a lesion charac-
Here we will focus our interest on the properties of terized by concentric rings of subnormal (< 1 mm),
single cells and the functional reorganization fol- hypernormal (1–2.5 mm), and normal spontaneous
lowing a cortical lesion that can again be related to activity (> 2:5 mm). The region of hyperactivity was
cellular equivalents of perceptual learning in the also characterized by epileptiform discharge patterns
primary visual cortex. in about one-third of the cells and by a typical loss of
orientation specificity to visual stimulation. Although
3.4.1 Acute Lesion-Induced Changes in the early hyperexcitability of cells surrounding a le-
Excitability sion could lead to excitotoxic cell death, we have
reasons to hypothesize that it could just as well sup-
Acute and subacute e¤ects were observed at small, port synaptic plasticity and reorganization by facili-
round lesions of about 1.5 to 2 mm diameter in- tating heterosynaptic LTP-like mechanisms.
duced either by surface photocoagulation (Eysel and
Plasticity of Receptive Fields of the Adult Visual System 53
Figure 3.7
Significantly enlarged receptive fields (RFs) at the border of a chronic visual cortex lesion after two months. (A) Surface map
with cortical vessels drawn from a macroscopic photograph. The positions of electrode penetrations (1p–4p) are indicated by
crosses ( prelesion recordings 1–4) and black circles (76-day postlesion recordings 1–4). (B) Control mapping at electrode
positions 1 to 4 before lesioning. Anteroposterior positions of recording sites in the right hemisphere are shown schemati-
cally above and the location of the injection of 2% ibotenic acid (IBO; 500 nL). (C) Same cortical region is remapped after
76 days (76d). The electrode tracks (1p–4p) are shown above, the RF sizes below indicate a significant increase of RF sizes at
the border of the lesion ( penetrations 2p and 3p). (Adapted from Eysel and Schweigart 1999.)
ing and final experiment. The localization of pene- cells could be determined with the same accuracy as
trations was documented by photographing the cor- the RFs during the initial lesioning experiment in
tical surface on the day of ibotenic acid injection, the same animal. At recording sites 2p, 3p, and 4p,
enabling us to retest the same cortical positions after close to the posterior and anterior border of the le-
the 76-day survival time: the recording sites and the sion, individual cells displayed impressingly increased
lesion site were characterized by certain landmarks in RF sizes (figure 3.7C), although in the same pene-
the vascular pattern on the cortical surface (figure trations we also encountered single cells with rather
3.7A). Panel B shows the RF locations and sizes as normal RF size (the smallest and largest fields obtained
mapped before lesioning. Because the cells were in penetration 2p). RFs showed normal ( prelesion)
vigorously responding, the location and outermost sizes at the most posterior recording site (1p) and
borders of the excitatory RFs could be exactly map- again decreasing RF size at the most anterior
ped. The largest RF found in each penetration is recording site (4p). This finding does not reflect the
outlined and labeled with the number of the respec- normal situation in the visual cortex where RF size
tive electrode penetration. The ibotenic acid injec- increases with distance from the area centralis, and
tion was applied between the locations of the hence a continuous increase in average RF size
recording sites 2 and 3 (figure 3.7A,B). After the 76- should be present from the most posterior to the
day survival time, we recorded single cells with crisp most anterior penetration. The cells with increased
responses, direction specificity, and sharp tuning for receptive field size were situated close to the border
orientation in the supragranular cortical layers in of the chronic lesions as evident from the position of
penetrations anterior or posterior to the lesion site. the recording sites relative to the ibotenic acid in-
The location and spatial extent of the RFs of these jection site (figure 3.7A). This close vicinity was
56 Ulf T. Eysel
histologically verified in sections that both showed size and thus to an enlargement of the visual field in
the lesion and identified the electrode tracks. most patients. The recovery was specifically depen-
With an array of seven electrodes spaced 1 mm dent on practice, which was also the case in a study
apart in the visual cortex and left in place from be- where patients with homonymous visual field defi-
fore to two days after the lesion, the mean size of 15 cits were exposed to computer-based visual field
RFs stayed unchanged anterior and posterior to the training that specifically activated the border region
lesion (mean: 101%; range: 75–156%). The scatter of of the visual field defect (Kasten and Sabel 1995;
absolute RF sizes also remained unchanged as can be Kasten et al. 1998). It was hypothesized that this re-
seen in values before/after lesioning (mean: 3.6 / covery might take place at the level of the striate
3.5 ; median: 3.2 /3.2 ; range: 2–6.4 /2–5.6 ; stan- cortex. The enlarged RFs of cells at the border of the
dard deviation: 1.32 /1.06 ), indicating, on average, cortical lesion represent a mechanism for functional
no change in size within the first two days after recovery of primarily lost parts of the visual field and
lesioning. When RF sizes before lesioning ðn ¼ 8Þ thus lead to a reduction of the size of a scotoma. This
were compared with those at the same topography is a nonredundant, functionally useful reorganization
in the visual field 55 and 76 days after lesioning because inputs that have lost their target cells are
ðn ¼ 18Þ, the mean RF size was found to have in- reconnected to surviving cortical cells. The absolute
creased to 182% (range: 64–385%). The larger scat- increase of RF size (figure 3.7C) would allow for a
ter and increased mean RF size were due to the cell shift of the border of a scotoma by about 3–4 in the
RFs within a distance of about 1 mm from the bor- cat, which is in the range of the 4.9 G 1.7 recently
der of the lesion. Although RF width remained observed in patients treated with computer-based
constant (close to 100% its size before lesioning) at visual field training (Kasten et al. 1998). This find-
1 mm anterior as well as 1 and 2 mm posterior, there ing suggests that the long-term plasticity of cells at
was a mean increase to 188% at the posterior border, the border of visual cortical lesions may represent a
and to 182% and 282%, respectively, at the anterior model in the mature visual system of the cat for the
border and 0.5 mm away. Though close to the le- long-term reduction of visual field defects observed
sion, 13 out of 14 RFs were larger than the mean of in human patients.
all prelesion RFs by up to 7.8 ; the RFs recorded 1– In light of determining that receptive fields can be
2 mm away were exactly in the range of the average changed in size and spatial substructure by repetitive
prelesion RF sizes (0.8 to þ0.2 ). visual stimulation, the next step was to test whether
RFs in the cat cortical lesion model could be made
Reducing the Size of the Scotoma to extend into the previously lost part of the visual
Patients with visual field loss due to vascular or field by visual training.
traumatic postgeniculate damage with homonymous
defects were trained in the border regions of their
3.5 Fast Lesion- and Training-Induced
residual visual fields by repeated stimulation to im-
Changes in Receptive Field Size near Subacute
prove light di¤erence thresholds (Zihl and von Cra-
Lesions in Adult Cat Visual Cortex
mon 1979) or by locating targets within the blind
field region (Zihl and von Cramon 1985). Both
Most studies of plasticity in sensory cortices have
types of training led to a reduction of the scotoma
concentrated on the remarkable changes of the cor-
Plasticity of Receptive Fields of the Adult Visual System 57
tical topography (somatotopic or retinotopic orga- 3.5.1 Lesion-Induced Changes in Receptive Field
nization, respectively) or on the extent of behavioral Size
recovery (for example, by measuring the size of a
scotoma). Comparatively little is known, however, Receptive fields were quantitatively mapped before
of the extent to which individual neurons may lesioning and their topography and size were deter-
change their response properties in the vicinity of mined (Schweigart and Eysel forthcoming). From
a cortical lesion. Jenkins and Merzenich (1987) re- the most posterior to the most anterior electrode of
ported a complete filling in of the acutely lost rep- the array in the visual cortex, the RF position of
resentation of the palm after an experimental lesion the cells migrated from the central visual field
in the primary somatosensory cortex of an adult downward, and the average RF size increased with
owl monkey. This functional reorganization was due eccentricity.
to strongly enlarged receptive fields of the cells sur- RF sizes were, on average, no di¤erent before/
rounding the lesion after 129 days, a finding very after lesion (mean: 3.1 /3.0 ; median: 2.8 /2.8 ;
similar to the increased RF sizes we observed in range: 0.8–10.4 /0.4–7.2 ). Although both the width
Eysel and Schweigart 1999 at the border of lesions in and length of the receptive fields remained constant at
the adult cat visual cortex after two months, but not most positions, data scatter was usually increased close
during the first two days of survival time. to the lesions (aG1 mm), and RF widths close to the
In Schweigart and Eysel forthcoming, we con- anterior border of the lesions (within 1 mm) were
centrated on the first two-day period after the lesion. slightly enlarged (width increase: 0.4–1.2 , p < 0:05;
We recorded neurons in the most interesting border length increase: 0–2.4 , p < 0:1; paired one-tailed t-
zone of 1–2 mm around the lesion to evaluate their test). There was a significant positive correlation be-
RF sizes and their ‘‘on’’ and ‘‘o¤ ’’ subfields and to tween change in neuronal activity and change in RF
determine whether use is a necessary prerequisite to size.
induce the RF size increase early after a lesion. Thus In summary, we observed modest RF size changes
we kept the animals in total darkness for two days in the immediate vicinity of the lesion, which cor-
after lesioning before we applied a visual training relate to lesion-induced changes in activity. Activity
procedure. The ibotenic acid lesions closely resem- changes were inversely correlated to the distance
bled those in the above-described chronic phase from the lesion. RF changes were not found to de-
study: average lesion diameters were 2.9 mm pend on other parameters; in particular, they did not
anteroposterior (median: 2.5 mm; range: 2.1–5.4 depend on horizontal or vertical eccentricity, RF
mm); 2.8 mm mediolateral (median: 2.7 mm; range: type (simple or complex), and cortical area (area 17
1.7–4.2 mm), and 2.0 mm dorsoventral (median: or 18).
1.7 mm; range: 1.2–3.6 mm). Histology supported
the electrophysiological finding that as of postlesion 3.5.2 E¤ect of Visual Stimulation on Receptive
day 2 no neurons had survived in the core of the Field Size
lesion, although neurons that could be recorded
in the surround were anatomically and functionally Two days after lesion, we tried to make single cells
intact. expand their RF width as a result of visual training.
After visual costimulation (see above), the neurons
58 Ulf T. Eysel
could be a¤ected both in an unspecific or specific lated side of the RF and no expansion to the other,
way. Unspecific e¤ects were characterized by an in- nonstimulated side. RF expansions were only mod-
crease in RF size directed toward both sides, the est (on the order of 1 ); the e¤ects could be seen
costimulated and the opposite, nonstimulated side, after 20–60 min of visual costimulation, and the
together with an increase in spontaneous activity and induced e¤ect lasted for approximately one hour.
visually evoked responses within the RFs. Where a Visual costimulation often induced, apart from a
RF expansion was induced in the formerly unre- specific RF enlargement, a slight general increase
sponsive region at the costimulated side of the RF, in activity (both spontaneous and visually evoked
but not at the nonstimulated side, a specific increase activity) in the lesioned and in the normal cortex.
of RF width was observed in roughly half of the Thus early visual training of cells close to the border
tested neurons (6/10). Figure 3.8A,B shows a typical of a cortical lesion can e¤ectively be applied to in-
example, and figure 3.8C displays the statistical eval- crease RF sizes of single cells by about 1 within
uation of the results from the neurons with specific minutes to one hour of training and can thereby re-
e¤ects. For statistical evaluation, we concentrated on duce the size of the scotoma; that is, the input to an
five di¤erent positions: at the RF border of the non- enlarged part of the RF comes from retinal regions
stimulated side, (a) just outside the RF and (b) just originally within the scotoma. Thus also formerly
inside the RF; on the stimulated side, (c) in the RF subthreshold synapses of the synaptic ‘‘integration
center and, at the RF border, (d) just inside the RF, field’’ (Pei et al. 1994; Frégnac et al. 1996) may be
and (e) just outside the RF. We made three observa- activated by a LTP-like mechanism (Eysel, Eyding,
tions. The values at the nonstimulated RF border, (a) and Schweigart 1998).
and (b) in panel C, did not di¤er before and after
visual costimulation. Moreover, the activity just
outside the RF did not di¤er from the spontaneous 3.6 Mechanisms for Modifying Visual
activity after costimulation; thus the RF did not ex- Receptive Fields
pand toward this side. At the costimulated sites,
visual responses were increased at both the RF cen- 3.6.1 Changes in Activity and Imbalance of
ter, though the di¤erence was not significant, and the Transmitter Systems
stimulated RF border, see the position just inside the
RF, (d) in panel C. In addition, after costimulation, Several studies have shown an increase in activity at
the RF was expanded at the costimulated border; the the surviving border of neocortical lesions. Heat
activity just outside of the RF was significantly dif- lesions in cat visual cortex yielded subnormal activity
ferent from the spontaneous activity ð p < 0:05Þ; see in this region of less than 1 mm around the lesion,
(c) in panel C. RF increases ranged from 0.4 to 0.8 whereas hyperactivity was only seen at 1–2.5 mm
from the border of the lesion; at more distant posi-
(although the visual costimulus often extended far-
tions (> 2:5 mm) the activity was normal (Eysel and
ther than 0.8 across the RF boundary).
Schmidt-Kastner 1991; for reviews, see Eysel 1997;
The results compared well with those obtained
Eysel et al. 1999). This hyperactivity after heat
from normal cats (Eysel, Eyding, and Schweigart
lesions was already present one day after the lesion
1998; see also above). Approximately half of the
and was still visible (although no longer significant)
neurons showed specific expansion to the costimu-
Figure 3.8
Training for receptive field (RF) size increment at the border of a 1-day (1d) visual cortical lesion. (A) RF size of a cell at the
border of a cortical lesion after 2-day (2d) survival time (thick solid line). RF size comprises ‘‘on’’ and ‘‘o¤ ’’ response regions
as obtained with the objective reverse correlation method. The cat was kept in darkness after lesioning. After the initial test,
costimulation with a large field ‘‘on-o¤ ’’ stimulus (white square surrounded by dashed line) was applied for 60 min. (B)
Resulting expansion of the RF is shown by the new solid line field boundaries as compared to the previous RF size (dotted
line). Right visual field, with elevation (vertical position) plotted on the ordinate; horizontal position in degrees of visual
angle, on the abscissa. The area of the scotoma is shown in darker gray. (C) In about half of the neurons investigated with
the method described in panels A and B a modest but significant ( p < 0:05) and specific RF increase was observed. Note
the significant increases of normalized activity at positions (d) and (e) on the RF side where the visual costimulation was
applied. The activity after costimulation (gray columns) was significantly higher than the activity before costimulation (white
columns). The average visual evoked activity ( peak minus background activity) is normalized with respect to the maximal
response in the RF center before costimulation (¼ 1) and shown for five di¤erent RF positions.
60 Ulf T. Eysel
30 days after lesioning. Similar results were obtained were recorded intracellularly with sharp electrodes,
in experiments photochemically inducing thrombosis revealing EPSPs with longer durations and higher
in visual cortex (Eysel, Kretschmann, and Schmidt- amplitudes in regions having increased field poten-
Kastner 1993) and in somatosensory cortex (Domann tials. The upregulation of the glutamatergic responses
et al. 1993; Schiene et al. 1996). Here mean discharge was accompanied by a reduced GABA-mediated in-
frequency was slightly higher in the immediate sur- hibition 1–5 days after lesion and 1–2 mm from the
round of the lesion (approximately 1 mm from the border of the lesions. Both the fast GABAA -induced
lesion border), but became considerably higher at a and the late GABA B -induced IPSPs showed reduced
distance of 2–3.5 mm, whereas at distances greater amplitudes and peak currents. Although no sponta-
than 4 mm lateral to the lesion border, the electro- neously occurring bursts were observed, epileptiform
physiological responses did not significantly di¤er burst activity could be evoked by strong stimuli ap-
from those of the controls. These alterations began plied to the white matter. Interestingly, after retinal
on the first day after lesioning and were most pro- lesions a quite similar hyperactive zone was found in
nounced 3–7 days after lesioning. the visual cortex at the border of the resulting corti-
One possible mechanism for the increase of activity cal scotoma, where a simultaneous increase in gluta-
in the surround of lesions is an imbalance between mate immunoreactivity was observed (Arckens et al.
GABAergic inhibition (loss) and glutamatergic exci- 2000). On the other hand, downregulation of the
tation (increase). Mittmann et al. (1994) found that GABAergic system within the cortical scotoma was
inhibitory postsynaptic potentials (IPSPs) were de- demonstrated by decreased GAD immunohisto-
creased; Schiene et al. (1996) found that GABA chemistry (Rosier et al. 1995) and decreased GABA
receptors were reduced in widespread brain areas levels (Arckens et al. 2000).
surrounding the lesion. Changes in glutamatergic The main lesion-induced e¤ect may be the above-
excitation were expressed in higher amplitudes and mentioned increase of excitability (as indicated by
longer durations of excitatory postsynaptic potentials increases in both spontaneous and visually evoked
(EPSPs) mediated by N-methyl-d-aspartate (NMDA) activity) in the surround of lesions; the changes in
receptors (Mittmann et al. 1994). The slice prepara- RF size may build on this changed excitability. A
tion allowed researchers to study the single-cell and correlation of an increase in activity with an increase
network properties in the vicinity of a focal lesion in in RF size in visual cortical neurons was found after
vitro. Field potentials were extracellularly recorded retinal lesions or when using the paradigm of the
in the perilesional area at survival times between 1 artificial scotoma (for review, see Gilbert 1998).
and 6 days after lesioning ( panel A of figure 3.9). When using bicuculline, a GABAergic antagonist
Although potentials very close to the lesion were that strongly enhances neuronal activity, RF expan-
strongly depressed, as observed in vivo, in the adja- sion was found in the somatosensory cortex of cat
cent region (1–1.5 mm from the border of the le- and rat (Dykes et al. 1984; Kyriazi et al. 1996). These
sion), the field potentials were significantly larger RF enlargements are discussed in terms of subthres-
than normal 1–5 days after lesioning. The enlarged hold inputs of intercolumnar connections that may
potentials were blocked with d-amino-phosphono- be enhanced by the bicuculline application. Pern-
valeric acid (APV), and thus are NMDA receptor berg, Jirmann, and Eysel (1998) found a widening of
dependent. The underlying single-cell responses RF subfields in cat visual cortex (A18) after local
Plasticity of Receptive Fields of the Adult Visual System 61
Figure 3.9
Increased field potentials and facilitated long-term potentiation in the surround of heat lesions in the rat cortex in vitro. (A)
Schematic drawing of the lesion with recording sites marked by black dots. The field potentials were recorded from the
cortex after electrical stimulation of the white matter. Note the depressed response very close to the border and the hyper-
excitability in the adjacent recording site 1 mm from the border of the lesion. (Modified from Mittmann et al. 1994.) (B)
Long-term potentiation (LTP) was induced in age-matched sham-operated rats (control) and rats with 1–6 days survival
time after infrared laser lesions of the visual cortex (lesion). Postsynaptic potentials were recorded intracellularly; typical
example and statistical evaluation for 6 control and 13 lesion experiments are shown. Excitatory postsynaptic potentials
(EPSPs) had about 2.5 mV amplitudes before theta burst stimulation. During the first six days after lesioning, LTP was much
stronger in the surround of lesions (2–4 mm) than in control animals (190.4% versus 136.7%). (Adapted from Mittmann and
Eysel 1999.)
application of bicuculline. Sober et al. (1997) found 3.6.2 Growth Factors and Morphological Correlates
that, after subacute lesions, an increase in activity is of Receptive Field Plasticity
paralleled by an increase in RF size in macaque
middle temporal cortex (area MT, multiunit record- The neurotrophins brain-derived neurotrophic factor
ings). The factor of RF expansion was 1.1–7.0. As in (BDNF), neurotrophin-3 (NT-3), nerve growth fac-
our work described here, the RF expansion of Sober tor (NGF), and insulin-like growth factor 1 (IGF-1)
et al. (1997) was not consistently directed toward the have all been found to be elevated in visual cortex as
lesion. Because of the observed correlations between early as 3 days after binocular retinal lesions (Obata
RF size and excitability in these studies and our own et al. 1999); the related neurotrophin receptors were
recent study (Schweigart and Eysel forthcoming), we elevated as well. This finding together with that of
suggest that the early increase in RF size is at least increased transcription levels of calcium calmodulin–
partially due to the increased excitability in individual dependent kinase II (CaMKII), microtubuli-asso-
neurons and to already existing but primarily sub- ciated protein 2 (MAP-2), and synapsins in visual
threshold synapses rising above threshold. cortex after such lesions provide evidence for addi-
62 Ulf T. Eysel
tional factors associated with cortical reorganization. Eysel, and Machemer 1991; Crook and Eysel 1992;
The increased level of BDNF in turn may be related Crook, Kisvárday and Eysel 1996) and lesions (Eysel
to the increased activity in regions with elevated and Schmidt-Kastner 1991; Domann et al. 1993;
BDNF expression (Prakash, Cohen-Cory, and Fros- Mittmann et al. 1994) can reduce lateral inhibition.
tig 1996) and hence may combine with GABAergic A ring of increased excitability (high spontaneous
and glutamatergic mechanisms to trigger the early activity, increased visually evoked activity) around
unmasking of existing connections. Furthermore, the cortical lesion and the cortical region dea¤erented
BDNF may also represent a link to the late mor- by a retinal lesion during the first days to weeks of
phological changes involving axonal sprouting and survival time appears to occur in parallel to or as
synaptogenesis (Gilbert and Wiesel 1992; Darian- a consequence of increased extracellular glutamate
Smith and Gilbert 1994). levels in the surround of cortical lesions (Choi and
Rothman 1990) and increased glutamate immuno-
3.6.3 Long-Term-Potentiation-like E¤ects and the reactivity in cells at the border of the dea¤erented
Role of Learning and Training in the Visual Cortex cortex after retinal lesions (Arckens et al. 2000). A
similar hyperexcitability was also present in slice
The finding of enlarged receptive fields at the border preparations from animals with 1–6 days of survival
of cortical lesions raised the question of how and time after cortical lesions in vitro. Under this condi-
under what special conditions cells can increase their tion, the NMDA receptor–mediated EPSPs were
receptive field size in the adult, completely wired increased in the same region that was characterized
visual cortex. We tested whether heterosynaptic by increased excitability in vivo (Mittmann et al.
LTP-like mechanisms might be a prime candidate 1994). In addition, the fast GABAA -mediated and
for such RF changes (cf. chapter 7.1). Applying re- the slower GABAB -mediated IPSPs were strongly
petitive synchronous stimulation to the receptive reduced (Mittmann et al. 1994) and GABAA recep-
field center of simple cells and regions just outside tors were downregulated (Schiene et al. 1996) in the
their receptive fields for 10 to 60 min, in many cases region surrounding a cortical lesion. Related obser-
we were able to extend the area of the excitatory vations were the significantly reduced GAD immu-
receptive field of the investigated cells specifically noreactivity in the neuropil (Rosier et al. 1995) and
into the coactivated region (Eysel, Eyding, and the decreased GABA levels (Arckens et al. 2000)
Schweigart 1998). This e¤ect showed a long-term within the cortical scotoma resulting from retinal
recovery over hours. We interpreted this as a hetero- lesions.
synaptic LTP-like increase of the e‰cacy of for- The evidence to date indicates that an early shift
merly subthreshold inputs from the region just toward locally increased excitability ( both due to
outside the classical excitatory receptive field. Because, increased NMDA-mediated and decreased GABA-
according to in vitro observations (Hirsch and Gil- mediated transmission) accompanies or precedes
bert 1993), lateral inhibition can suppress such LTP- neuronal reorganization in response to two quite
like e¤ects in cat visual cortex, reduction of lateral di¤erent types of damage in the visual system: (1)
inhibition should facilitate LTP-like mechanisms. As retinal lesions in the eye and (2) cortical lesions in
mentioned above, both intracortical local inactiva- the brain. Transient local reactions of the inhibitory
tion (Eysel, Crook, and Machemer 1990; Crook, and excitatory cortical systems on the synaptic level
Plasticity of Receptive Fields of the Adult Visual System 63
are initially observed at the sites of lesion-induced Parallel to changes in the retinotopic maps, the sizes
reorganization irrespective of the kind of lesion. of receptive fields in the visual cortex have also
Increased excitability due to NMDA-mediated de- shown long-term and short-term changes in re-
polarization may be the most important trigger sponse to retinal lesions (Gilbert and Wiesel 1992;
for synaptic plasticity. The reduced inhibition may for reviews, see Chino 1997; Gilbert 1998) or arti-
play an additional important role because reduced ficial scotomas (Pettet and Gilbert 1992; Das and
GABAergic inhibition can lead to increased activa- Gilbert 1995b). Moreover, RF expansion has been
tion of NMDA receptors (Luhmann and Prince, found after repetitive visual stimulation with stimuli
1991). Finally, the reduced activity in cells that are covering the RF center and an unresponsive region
targets of synaptic reorganization may reduce the outside the RF (Eysel, Eyding, and Schweigart 1998):
threshold for successful potentiation (‘‘sliding thresh- the stimulus-induced RF expansion occurred in the
old model’’; Bienenstock et al. 1982). Based on the region outside the RF that was costimulated with the
above-mentioned observations, we have predicted RF center. At the border of focal excitotoxic lesions
(Eysel 1997) the facilitation of LTP in the vicinity of in the visual cortex of cats, the same repetitive stim-
visual cortical lesions. ulation protocol induced similar increases of RF size
Two mechanisms seem to prevail in lesion- in the subacute phase (Schweigart and Eysel forth-
induced and use-dependent reorganization of RF coming).
properties: a short-term mechanism acting within Perceptual learning can play a key role in adult
minutes to hours and a long-term mechanism taking visual RF plasticity; LTP-like mechanisms are prime
weeks to months to develop (see Gilbert 1998 for candidates for the underlying modifications of syn-
review). Reorganization of the long-term mecha- aptic e‰cacy. Homosynaptic LTP and heterosynaptic
nism, attributed to local sprouting (growth of axonal LTP have been documented in vitro in slices of the
collaterals and synaptogenesis) of the long-range visual cortex of adult rat (Artola and Singer 1987;
horizontal connections within the cortex (Darian- Kossel, Bonhoe¤er, and Bolz 1990; Kirkwood and
Smith and Gilbert 1994, 1995), ranges over some Bear 1994) and cat (Hirsch and Gilbert 1993). Many
6–8 mm in the visual cortex, although alterna- in vivo studies have shown LTP-like changes of spe-
tive pathways involving feedback projections from cific RF properties when repetitive stimulation was
higher-order cortical areas have not yet been ruled paired with postsynaptic depolarization in the visual
out. In contrast, the short-term mechanism produces cortex (Frégnac et al. 1988, 1992; Frégnac and Shulz
shifts of the cortical retinotopy within minutes to 1999). Associative (heterosynaptic) LTP might also
hours, ranging over only some 2 mm in the cortex. be the basis of the long-term e¤ects elicited by syn-
Because they are in the range of the spread of thala- chronous costimulation of a nonresponsive RF re-
mocortical a¤erents and are fast, these shifts were gion together with the excitatory center of the RF
attributed to the unmasking of formerly subthreshold in the adult cat visual cortex in vivo (Eysel, Eyding,
thalamocortical synaptic inputs rather than to the and Schweigart 1998). Intracortical lateral inhibition
forming of new connections (Gilbert 1998). In fact, can suppress LTP in vitro (Hirsch and Gilbert 1993)
the EPSP field area is about nine times larger than and can also prevent receptive field plasticity in vivo,
the spike discharge field (Pei et al. 1994; see also the thus keeping the RF size and substructure stable and
synaptic ‘‘integration field’’ of Frégnac et al. 1996). the subfields separated unless inhibition is reduced.
64 Ulf T. Eysel
This was directly shown by the increase of RF size Schweigart 1999), indicating that mechanisms of
and loss of simple cell RF substructures when long-term recovery span larger distances and inte-
GABAergic inhibition was microiontophoretically grate across a larger cortical representation than do
blocked with bicuculline (Pernberg, Jirmann, and those of short-term recovery. The larger changes in
Eysel 1998). A similar shaping of RFs by inhibition RF size as observed in chronic phase cats seem to
was also shown by application of bicuculline in the need longer training, longer time periods (over weeks
primary somatosensory cortex of the cat (Dykes et al. to months), or both, and may additionally involve
1984). Changes in the balance between excitation anatomical changes, such as axonal sprouting. This
and inhibition were also held responsible for the finding indicates that short-term recovery, even
fast changes in RF size observed after lesions in area when using visual training, seems not su‰cient for
MT of the macaque monkey (Sober et al. 1997). an extended recovery of lost visual function and that
Intracortical local inactivation (Eysel, Muche, and additional long-term mechanisms have to be e¤ec-
Wörgötter 1988; Eysel, Crook, and Machemer tive over longer survival times. After retinal lesions,
1990; Crook, Kisvárday, and Eysel 1998) and le- when the primary cortical reorganization has taken
sions (Eysel, Wörgötter, and Pape 1987; Eysel and place on the basis of preexisting connections and
Schmidt-Kastner 1991; Domann et al. 1993; Mitt- functional strengthening of synaptic weights (Das
mann et al. 1994; Schiene et al. 1996) seem to re- and Gilbert 1995a,b), the functionally modified
duce lateral inhibition. This reduction of lateral synaptic connections can become stabilized by ter-
inhibition could facilitate LTP and LTP-like mech- minal sprouting, as observed more than six months
anisms. Lesions in the somatosensory cortex have after retinal lesions in adult cat visual cortex (Darian-
been shown to induce LTP (Hagemann et al. 1998). Smith and Gilbert 1994). Because these late e¤ects
In Mittmann and Eysel 2001, we induced LTP in after retinal lesions are preceded by early changes in
vitro with repetitive stimulation of synaptic inputs RF size (Gilbert and Wiesel 1992), by a down-
combined with postsynaptic depolarization and have regulation of the GABAergic system (Rosier et al.
compared EPSPs of single cells in the normal adult 1995), and by an upregulation of the glutamatergic
rat visual cortex with those from age-matched rat system (Arckens et al. 2000), as observed at the bor-
cortices with focal infrared laser lesions having sur- der of cortical lesions (Mittmann et al. 1994), one
vival times of 1–6 days, and indeed found a sig- can expect a similar long-term plasticity at the bor-
nificantly enhanced LTP in the surround of focal der of chronic lesions in adult cat visual cortex.
cortical lesions. Panel B of figure 3.9 exemplifies the The above-mentioned psychophysical observa-
nearly threefold facilitation of LTP in the surround tions in human patients indicate that a scotoma
of cortical lesions 1–6 days after lesioning. LTP resulting from central ( postchiasmatic or post-
amounted to more than 190% in the lesioned ani- geniculate) brain damage can be reduced in size by
mals, whereas it amounted to less than 137% in the giving patients visual training in the border regions
sham-operated controls. of their residual visual fields (Zihl and von Cramon
On the other hand, increases in RF size were 1979) or by locating targets within the blind field
more pronounced after two months (Eysel and region (Zihl and von Cramon 1985; Kasten and
Schweigart 1999) than after short-term visual train- Sabel 1995; Kasten et al. 1998). It was suggested
ing (Eysel, Eyding, and Schweigart 1998; Eysel and that a correlate of this recovery of previously blind
Plasticity of Receptive Fields of the Adult Visual System 65
Figure 4.1
Lateral (top) and ventral (bottom) view of cortex, showing
the visual cortical areas along the occipitotemporal path-
way. The extent of TE is shaded.
Figure 4.5
One TE cell responding to the critical feature in a wide range of size. Horizontal bars below histograms indicate the duration
of the stimulus presentation; numbers to the right, the magnitude of the response normalized by that of the maximum re-
sponse. (From Tanaka et al. 1991.)
Figure 4.7
One TE cell selective for the orientation of the critical feature. Horizontal bars below histograms indicate the duration of the
stimulus presentation; numbers to the right, the magnitude of the maximum response.
determined the critical feature using the reduction of the first cell. The selectivities of the two cells dif-
method, and then tested the e¤ects of changes in the fered slightly, however, in that the maximal response
aspect ratio of the critical feature. We observed that was evoked by slightly di¤erent stimuli, or the mode
51% of cells responded to an aspect ratio range of of the decrease in response was di¤erent when the
more than three octaves with more than 50% of stimulus was changed from the optimal stimulus.
the maximum responses (Esteky and Tanaka 1998). To determine the spatial extent of the clustering
Figure 4.6 shows an example of such cells. of cells with similar selectivities, we examined the
Responses of TE cells are more selective for the responses of cells successively recorded along the
orientation of stimuli in the frontoparallel plane length of long penetrations made vertically or
(Tanaka et al. 1991). Figure 4.7 shows an example. obliquely to the cortical surface (Fujita et al. 1992).
Rotation of the critical feature by 90 decreased the The critical feature for a cell located at the middle of
response by more than 50% for most cells. the penetration was first determined. A set of stimuli,
including the critical feature for the first cell, its
rotated versions, and ine¤ective control stimuli, was
4.4 Columnar Organization in Area TE
constructed, and cells recorded at di¤erent positions
along the penetration were tested with the fixed set
We examined the spatial distribution of the cells
of stimuli. As in the example shown in figure 4.8,
responding to various critical features in area TE. By
cells recorded along the vertical penetrations com-
recording two TE cells simultaneously with a single monly responded to the critical feature for the first
electrode, we have found that cells located close
cell or some related stimuli. The span of the com-
together in the cortex have similar stimulus selectiv- monly responsive cells covered nearly the entire
ities (Fujita et al. 1992). The critical feature of one thickness from layers II to VI. The situation, how-
isolated cell was determined by using the same pro-
ever, was di¤erent in the penetrations made obliquely
cedure as described above, while the responses of
to the cortical surface. The cells that were commonly
another isolated cell, or nonisolated multiunits, were responsive to the critical feature of the first cell or
simultaneously recorded. In most cases, the second
related stimuli were limited to within a short span
cell responded to the optimal and suboptimal stimuli
Neuronal Representation of Object Images 73
Figure 4.9
Schematic drawing of the columnar organization in TE.
(From Tanaka 1996.)
Figure 4.12
Systematic movement of the activation spot with rotation of the face. The images were obtained for five di¤erent views of
the same doll face shown at top. The reference image obtained by averaging the five images has been subtracted. The con-
tours circumscribing the pixels with t-values at p < 0:05, compared with the reference image, are superimposed at the
bottom. (From Wang, Tanaka, and Tanifuji 1996.)
other than those with horizontal rotation are repre- This systematic arrangement of related columns
sented at di¤erent sites not covered by the recording can be used for various kinds of computation neces-
chamber in the experiments or (2) only the variations sary for object recognition. One simple possible
along the horizontal rotation are explicitly mapped computation is the generalization of activation by
along the cortical surface as the first principal com- the horizontal excitatory connections (cf. chapter 1)
ponent, and other variations are imbedded in over- to nearby columns representing related features,
lapping cell populations. which we might call ‘‘selective blurring of activa-
Although the data for the nonface features are tion.’’ Another is the mutual inhibition among the
more limited, I hypothesize that nonface and face nearby columns for winner-takes-all type selection.
features involve similar structures and propose a The continuous mapping of di¤erent views of
modified model of the columnar organization in TE faces cannot be generalized to nonface objects. Be-
in figure 4.13. The borders between neighboring cause the critical features for TE cells are only mod-
columns are not necessarily distinctive. Instead, mul- erately complex except in the case of faces, the
tiple columns that represent di¤erent but related image of a nonface object has to be represented by a
features partially overlap with one another and as a combination of activations at multiple cortical sites.
whole compose a larger-scale unit. At least in some Rotation of a nonface object causes shifts of activa-
cases, some parameter of the features is continuously tion at multiple cortical sites, each of which corre-
mapped along the cortical surface. sponds to the partial change of a feature. To uncover
Neuronal Representation of Object Images 77
Figure 4.14
Twenty-eight shapes used for the training (left) and in paradigm (right).
from three untrained control monkeys, we calcu- trol monkeys responded maximally to some of the
lated the ratio of the maximal response to the train- stimuli. These results indicate that the number of cells
ing stimuli to the cell’s overall maximal response and maximally responsive to training stimuli increased
compared its distribution between the two groups during the period of the discrimination training. Sakai
(figure 4.16). The x-axis in the figure is the ratio, and Miyashita (1991, 1994) and Logothetis, Pauls,
and the y-axis the proportion of cells. The top his- and Poggio (1995; see also Sheinberg and Logothetis,
togram shows the distribution among 131 cells chapter 6, this volume) had trained adult monkeys
recorded in the trained monkeys, and the bottom to discriminate among fractal patterns or wire-frame
histogram shows the distribution for the same num- objects and found that many TE cells responded to
ber of cells recorded from the three control mon- the learned stimuli after the training. A unique con-
keys. One on the x-axis signifies that the cell was tribution of our study is the demonstration that
maximally activated by some of the training stimuli, training increases the proportion of TE cells that re-
and ‘‘0’’ signifies that the cell was not activated at all spond to the learned stimuli, as measured against
by any of the training stimuli. Twenty-five percent untrained controls.
of the cells recorded from the trained monkeys The responses to the training stimuli, rather than
responded maximally to some of the training stimuli, being sharply tuned to a particular stimulus, were
whereas only 5% of the cells recorded from the con- distributed for several di¤erent training stimuli; in-
Neuronal Representation of Object Images 79
Figure 4.15
Responses of one TE cell to the image of the most e¤ective object stimulus (top) and its responses to the twenty-eight stimuli
used for the training. This cell was recorded from a monkey that had been trained with the twenty-eight stimuli. Statistically
significant responses ð p < 0:05Þ are labeled with their relative response magnitudes. Horizontal bars below histograms indi-
cate the duration of the stimulus presentation; numbers to the right, the magnitude of the response normalized by that of the
maximum response. (From Kobatake et al. 1998.)
deed, detectors of particular training stimuli did not taken together, they covered the feature space
appear. Such training e¤ects were also found in occupied by the set of training stimuli. Cells recorded
responses of the cells to the eight stimuli shown in from the two trained monkeys responded to these
panel A of figure 4.17. These stimuli, referred to as hidden stimuli very well. The histogram at the mid-
‘‘hidden stimuli,’’ were not used during the training, dle of panel B shows the distribution of the normal-
but were presented for the single-cell recordings ized magnitude of individual responses to the hidden
under anesthesia. Because they were composed of stimuli of TE cells recorded from the trained mon-
the same primitives as those of the training stimuli, keys. It was nearly the same as the distribution of the
80 Keiji Tanaka
Figure 4.16
Distribution of the normalized magnitude of the strongest
responses of individual cells to the training stimuli.
Responses from 131 TE cells recorded from the two
trained monkeys are shown at the top; those from 130 TE
cells recorded from the three control monkeys are shown
at bottom. The magnitude of the response was normalized
with respect to the maximal response of the cell (the larger
of the strongest responses of the cell to the reference object
stimuli and to the training stimuli). (From Kobatake et al.
1998.)
thus should involve modifications of the sensory their specificity. Whereas ‘‘higher-order’’ learning
representations in the brain. The neurophysiological generalizes to other tasks, some forms of perceptual
correlates of perceptual learning are to be found in learning are highly specific to the particular stimulus
that area of the brain where processing of a particular attributes used while training. Although these find-
stimulus attribute occurs. ings suggest an early cortical site of plasticity, where
This brings us to another consideration about the the physiological characteristics of the neurons and
neurophysiological correlates of perceptual learning: their spatial organization correspond to the psycho-
the link between neuronal activity and perception. physical characteristics of the learning e¤ect, we
The di‰cult task of understanding the neural mech- should be cautious: neurons in higher-order areas
anisms underlying improvements in sensory capacity might exhibit the characteristics of neurons in lower-
is compounded by our failure to understand exactly order areas as a consequence of training.
how the cortex processes information. We expect
to find changes in aspects or characteristics of the
5.1 Lesion- and Use-Dependent Plasticity
neuronal response responsible for coding a particular
stimulus attribute that a subject is learning about.
Experimental manipulations such as nerve crush or
But very often, it is not clear what in a neuron’s
transection, digit amputation, surgical syndactyly,
firing pattern determines the code for a particular
skin translocation, or dorsal rhizotomy, on the one
stimulus attribute. Is it the size of the neuron’s re-
hand, and intracortical microstimulation or enhanced
ceptive field (RF)? Is it the sharpness of tuning, the
use, on the other, have been very important to dem-
strength of firing (cf. chapters 4, 6), or perhaps
the variability of the neuron’s firing rate, or even the onstrate how primary cortical sensory areas remain
under the influence of peripheral sensory activity
latency of its response?
Traditionally, we start from the assumption that throughout life (see Dinse and Merzenich, chapter 2,
this volume). These studies demonstrated changes in
information is carried in the time-averaged firing
the organization of the topographic maps in primary
rate. The alternative to this rate code is the temporal
sensory cortex. Reorganizations were demonstrated
code, where temporal aspects such as interspike
not only for the somatosensory and barrel cortex,
intervals and occurrence of bursts, rather than total
but also for the auditory cortex and the visual cortex
spike count, carry the information.
(for review, see Buonomano and Merzenich 1998a).
The search for neurophysiological correlates of
The timescale over which these changes occurred
perceptual learning, especially when considering a
ranges from minutes to weeks. For example, in the
visual task, is made even more di‰cult by the un-
certainty of where the changes could occur. About visual cortex, within minutes after making a retinal
lesion, cortical RFs near the boundary of the lesion
50% of our brain takes care of how we see things,
expanded in size, and the area of cortex initially
and neurons in many areas, although specialized to a
silenced to stimulation of the lesioned area recovered
great extent, still show overlapping characteristics.
input from the surrounding area (see Eysel, chapter
Traditionally, the principal indicator for localiz-
3, this volume). These changes are most likely at-
ing the neurophysiological correlates of perceptual
tributable to long-range intrinsic horizontal con-
learning that involves changes in sensory capacity is
nections (Gilbert 1998, cf. also chapter 1).
Electrophysiological Correlates of Perceptual Learning 85
cortical area of representation was the only parame- vations obtained in the developing visual system.
ter that correlated with behavioral performance. The primary visual cortex is organized in vertical
The reorganization described here involved a di- columns, such as ocular dominance columns, which
mension already present at the sensory periphery: group together cells with a similar ocular dominance,
selectivity for pure tone frequency. Moreover, the and orientation columns, which group together cells
auditory cortex is organized in a tonotopic manner. with the same orientation preference (cf. chapter 1).
Studies using associative training had previously Although the role of activity in modulating the
shown a selective retuning at the frequency signifi- formation of ocular dominance columns has been
cant during training, a retuning that did not occur well established (Hubel and Wiesel 1962), its role in
after pseudoconditioning. Associating a simple tone constructing higher-order receptive field properties,
of a defined frequency with an aversive stimulus such as orientation selectivity, remains controversial.
(foot shock) causes cells in primary auditory cortex That some neurons seem to shift their orientation
to increase their response to tones of that frequency, preference toward the experienced orientation upon
even for cells whose pretraining best frequency is enhanced exposure to a particular orientation is
di¤erent from the conditioning frequency (Wein- supported by studies of kittens reared with a biased
berger 1995; most experiments performed in guinea orientation, such as those reared in striped cylinders
pig). Thus the cortical area responding to the trained (Blakemore and Cooper 1970; Stryker and Sherk
frequency was found to be larger than that respond- 1975; Sengpiel, Stawinski, and Bonhoe¤er 1999),
ing to other frequencies, a logical consequence of the those reared with goggles containing images of lines
neurons’ having shifted their tuning to the trained (or of a single orientation (Stryker et al. 1978), or those
CS) frequency. reared with strong cylindrical lenses that blur all but
In another study on auditory discrimination train- a narrow range of contour orientations (Freeman and
ing in the unanesthetized Mongolian gerbils, the Pettigrew 1973). On the other hand, cortical orien-
spectrotemporal characteristics of the responses of tation maps are remarkably stable during the matu-
primary auditory cortex were examined by Ohl and ration period of orientation tuning, indicating that
Scheich (1997), who found both response increases visual experience is certainly not the only determi-
and decreases, with the decreases occurring much nant of cortical orientation selectivity during devel-
earlier than the increases. Their findings suggest that opment (Frégnac and Imbert 1978; Godecke and
the temporal organization of the neuronal discharges Bonhoe¤er 1996).
is more important than the spike rates in looking As described above, after the so-called critical pe-
for the neuronal substrate of information processing riod, once considered as the only period of plasticity,
during learning. restricted to early postnatal development, there is
ample evidence for lesion-induced plasticity in the
adult primary visual cortex, as well. Retinal lesions
5.3 Experience-Dependent Plasticity and were found to lead to remapping of the cortical to-
Perceptual Learning in the Visual System: pography, with a shrinkage of the representation of
Electrophysiological Correlates the retinal lesion and an expansion of that of the
surrounding part of the retina (Kaas et al. 1990;
In the visual cortex, much research on plasticity, Chino et al. 1992; Gilbert and Wiesel 1992; see
mainly use-dependent plasticity, is based on obser- Eysel, chapter 3, this volume).
Electrophysiological Correlates of Perceptual Learning 87
Extensive use or experience also seems to shape The second study (Rainer and Miller 2000)
our perception: when alphanumeric characters that looked at the ability to recognize objects, and found
are commonly tilted clockwise, such as displayed a neuronal correlate in the prefrontal (PF) cortex,
by digital clocks, are presented with a clockwise tilt, which receives input from the inferotemporal (IT)
they are perceived as less tilted than the same char- cortex, and plays a role in behaviors that involve
acters horizontally inverted (Whitaker and McGraw discriminating and remembering visual stimuli. Al-
2000). This form of permanent adaptation must find though familiar objects were found to activate a
a neurophysiological correlate in the visual neurons smaller population of PF neurons than did novel
coding for orientation. objects, these fewer neurons were more narrowly
Cellular conditioning as a form of learning was tuned, and their object representation, more resistant
also used in the visual system to demonstrate cortical to the e¤ects of degradation after experience. Inter-
plasticity. For example, Frégnac et al. (1988) used estingly, the largest improvements in neuronal per-
extracellular recordings from orientation-selective formance with familiarity were evident at the same
cells in kittens and cats and paired iontophoretically stimulus levels for which there was also the largest
driven neuronal activity with presentation of bars of improvement in behavioral performance.
light of varying orientation. A significant proportion The link between psychophysics and neuro-
(32%) of the sampled neurons exhibited a shift in physiology has also been the main interest of my
orientation preference to the orientation that was coworkers and me. We have used orientation dis-
paired with neuronal activity. crimination as a probe of early perceptual learn-
Besides the study presented below, two other ing. Orientation is a simple stimulus attribute that is
recent studies of perceptual learning in the visual first and probably best coded for by neurons in pri-
system have linked behavioral improvement to an mary visual cortex (area V1). Our main focus has
improved neuronal performance. In the middle been to find the neuronal changes responsible for
temporal visual area (MT) and the medial superior better coding by the neuronal population and even-
temporal area (MST), neurons have large receptive tually for better behavioral performance. To study the
fields and are typically direction selective. The first electrophysiological correlates in the adult macaque
study (Zohary et al. 1994) found that improvement monkey, we have used the following protocol: a
in direction discrimination at one position in the monkey was intensively trained on an orientation
subject’s visual field transferred to another location, discrimination task. By recording from single units
indicating that the site of the learning e¤ect most while the monkey was awake and performing a fix-
likely is at the level of area MT or MST. Within one ation task, we could fully characterize the tuning
training session, the increase in perceptual sensitivity properties of di¤erent populations of neurons. The
was accompanied by an improvement in neuronal purpose was to unravel changes in the firing pattern
sensitivity that mirrored the perceptual e¤ect both of trained neurons that not only appear as a result of
in magnitude and in time course, although the neu- training but, more importantly, could be responsible
ronal improvement proved to be transient: during for the improvement in the sensory discrimination
long-term training, it was not consolidated into task.
long-term gains in neuronal sensitivity (Zohary and The behavioral correlates of this type of perceptual
Newsome 1994). learning were documented previously in humans
88 Aniek Schoups
Electrophysiological Correlates of Perceptual Learning 89
(Schoups, Vogels, and Orban 1995; Vogels and gest that the neuronal correlates for this type of per-
Orban 1985), and were then repeated in monkeys. ceptual learning are likely to be found early in the
Subjects were trained daily in identifying the oblique visual processing pathway. Indeed, nowhere in the
orientation of a small circular grating, always at the brain are receptive fields as small as they are in area
same position in their visual field. Only one orien- V1. Moreover, early in the visual processing path-
tation, tilted either clockwise or counterclockwise way, neurons with similar orientation preferences are
with respect to the oblique reference orientation, grouped together, segregated in space from neurons
was shown on each trial, and the monkey had to with an orthogonal orientation preference. Instances
indicate its decision by a saccade to a point above or of learning could thus be associated with neurons
below the stimulus position. The reference orienta- tuned to one orientation without development of
tion was never shown. We strongly emphasized that an equivalent association with neurons tuned to an
any improvement observed had to originate at the orthogonal orientation. Extremely well suited to
sensory level and could not be attributed to a change code orientation, neurons in area V1 show a typical
in strategy. The phase was randomized so that the orientation tuning curve, characterized by the neu-
only cue that could be used to solve the task was the rons’ preferred orientation and by the selectivity of
orientation of the bars in the grating. The monkeys, tuning or sharpness of the tuning curve (see figure
just like the human subjects, improved their perfor- 5.3).
mance, with a large improvement initially, and with This specificity for both position and orientation
smaller improvements as training continued. Im- provided us with an interesting internal control: in-
provement was most evident between sessions. The stead of comparing data from trained monkeys with
most interesting results were the specificity of the those from naıve monkeys, we could use di¤erent
learning e¤ect: the improvement was highly specific populations of neurons within the same monkey.
to the position of the stimulus during training. A Indeed, neurons that have their RF at an untrained
mere displacement of the stimulus after training to an position in the monkey’s visual field could serve as a
adjacent position caused a marked increase in thresh- naıve population of neurons. The trained population
old (figure 5.1). Similarly, no transfer was observed of neurons consisted of neurons that had their RF
between orientations (figure 5.2). On the other hand, at the trained positions. Within this population of
complete transfer was observed between the two neurons, we specifically looked for changes in their
eyes (cf. chapters 9.5, 14.3.1, but also 10.3.4, 11.6). responses to the trained orientation.
The specificity for the trained orientation and the We introduced a second control population of
highly precise specificity for position together sug- neurons. During training, a second stimulus appeared
Figure 5.1
Position specificity of perceptual learning in orientation discrimination tested in two human subjects. (A) Overview of the
di¤erent positions for the grating (2.5 diameter) in the subjects’ visual field. The black dot in the center represents the
subject’s fixation point. (B) Learning curves for subject A.S. Just-noticeable di¤erences ( JND) represent the subject’s
thresholds. Orientation discrimination was first tested and trained at a central position, then di¤erent positions at 5 eccen-
tricity were tested and further trained. r ¼ foveal position; þ ¼ position 1; G ¼ position 2; ¼ position 3; k ¼ position
4; x ¼ position 5; h ¼ position 6; 9 ¼ position 7. (C ) Learning curves for subject A.C. r ¼ foveal position;
k ¼ position 6; 9 ¼ position 2; ¼ position 4; d ¼ position 2b. (Schoups, Vogels, and Orban 1995, figure 4.)
90 Aniek Schoups
Figure 5.2
Specificity for stimulus position and orientation in a macaque monkey trained for orientation discrimination learning. Graph
( left) shows the just noticeable di¤erences ( JND) that the monkey could discriminate; shaded circles (upper right) indicate
stimulus positions and orientations of the grating tested. Orientation discrimination with the stimulus at position 1 was tested
first for both oblique orientations, then training was continued for the 45 only. When threshold was reached, performance
for the nontrained oblique was tested again and found to be worse than before training the 45 . Orientation discrimination
with the stimulus at positions 2 and 3 in the same quadrant as position 1 were tested only once, without further training.
Some transfer was found for these stimulus positions, probably due to the monkey’s fixation errors during training. Untrained
performance for the stimulus at positions 4, 5, and 6 in other quadrants were worse than for positions in the same quadrant.
at a di¤erent position in the monkey’s visual field. These findings were always tested behaviorally:
Its orientation was randomly picked from a narrow after the recordings, we tested the monkey’s behavior
range of orientations around the oblique reference for the stimulus at the trained position, for the trained
orientation, orthogonal to the trained orientation. as well as other orientations; we also tested the mon-
The monkey learned to ignore this stimulus, but at key’s performance to identify these orientations for
the same time, the neurons that responded to stimuli a stimulus shown at the two control positions, one
at that location were being exposed to the stimuli as completely naıve, and one passively stimulated. Only
many times as the trained neurons, only they did not for the trained orientation, at the trained position,
learn the task. did the monkey perform well; its threshold for other
Electrophysiological Correlates of Perceptual Learning 91
would be broader at the position in the other hemi- Vogels, and Orban 1998). At first sight, this result
field trained for this orientation. Thus again, for each seems counterintuitive. Indeed, why would less fir-
training, we had an internal control. Another control ing result in better performance in a sensory dis-
came from the orientation columns at positions away crimination task?
from the trained ones; these were naıve populations Another change we observed in the population of
of neurons and should not have shown a broadening trained neurons was that some neurons, tuned to
or narrowing of the orientation columns. The results orientations 30 away from the trained orientation,
from this experiment, however, confirmed the elec- were broader than the analogous neurons from the
trophysiological data: no broadening of the orienta- naıve population. Theoretically, this could represent
tion columns was observed as a consequence of the a recruitment of neurons responding to the trained
perceptual learning, and thus no recruitment of cells orientation. For example, as shown in panel B of
responding to the trained orientation. figure 5.3, if neurons b and c broaden their tuning
At this point, I would like to go back to the curve, then the trained orientation, previously not
question what neuronal changes would be relevant. included under the curve, now does evoke a re-
How do the responses of primary visual neurons re- sponse. The response of most of these neurons is
flect orientation di¤erences? Even in area V1, orien- lower than to some other behaviorally less relevant
tation tuning curves are relatively broad, that is, 10 orientations. But the interest comes from the number
to 30 , with a large variability. One possible strategy of neurons responding. For a better coding, it may be
could be that the experienced cortex has learned to more interesting that more neurons respond, all at
select those neurons with the sharpest tuning curves, the same, low firing rate, than that fewer neurons
but even then, orientation di¤erences of only 0.5 respond, with a high interneuron variability.
cannot be discriminated using the response of the Building a model on perceptual learning in ori-
most excited neurons only. Instead, as demonstrated entation discrimination, Qian and Matthews (1999)
in panel A of figure 5.3, using the di¤erence signals found that a lower firing rate by the neurons that
between neurons that are tuned to orientations on prefer the trained orientation could lead to selective
either side of the trained orientation gives a much changes in the tuning patterns of neurons that prefer
higher sensitivity to detect orientation di¤erences the orientations bordering the one trained, which
(Westheimer, Shimamura, and McKee 1976; Regan then would lead to a better performance in the dis-
and Beverley 1985). Thus the changes we really crimination task.
should be looking for are changes in the slope of the
orientation tuning curves of the neurons that are
tuned to orientations, close to and on either side of 5.4 Human Learning and Imaging Studies
the trained orientation (Schoups et al. 2001).
Of the increases and decreases reported in imaging
When we examined the firing characteristics of
studies after learning a variety of tasks, I will especially
the trained neurons obtained from the single-unit
focus on those reported in the visual system. First, a
recordings, one change stood out sharply: in the
decrease is often associated with the task having be-
population of trained neurons, those preferring the
come automatic. The circuitry used when perform-
trained orientation exhibited a lower firing rate than
ing a task can change as a consequence of practice
the neurons preferring other orientations (Schoups,
Electrophysiological Correlates of Perceptual Learning 93
(see also Walsh, Ashbridge, and Cowey 1998). Sec- Karni and Sagi (1991) had reported learning to be
ond, a decrease in the neuroimaging signal with specific for the target position in the subject’s visual
learning may reflect learning-dependent improve- field, for the orientation, and for the eye (monocu-
ments in neural coding e‰ciency (Földiák 1990; lar). Our group (Schoups and Orban 1996) disputed
Barlow 1994; Scannell and Young 1999). It could the monocularity, however; it was di‰cult to imag-
also represent a more e‰cient ‘‘readout’’ of the in- ine that learning could be both orientation and eye
formation encoded at the sensory level. specific because monocular cells in striate cortex
Using positron-emission tomography (PET) to lack orientation specificity (cf., however, chapter 11)
study activation of brain areas after perceptual learn- Finally, three research groups have reported that the
ing in a visual discrimination task, researchers have fusiform gyrus shows an increase in activation when
reported large decreases in the visual areas. For ex- expertise is acquired in recognizing novel objects
ample, regional cerebral blood flow (rCBF) was (Gauthier et al. 1999) or degraded faces (Dolan et al.
measured twice in humans: before and after training 1997; Tovee, Rolls, and Ramachandran 1996).
orientation discrimination. In accord with our elec- The reported changes in rCBF with practice on a
trophysiology findings, significant and orientation- variety of tasks should be viewed with caution: they
specific decreases in activation were observed in striate may reflect an automization of the task, a change in
and extrastriate visual cortex (Schiltz et al. 1999). Also the involvement of attention, a familiarity e¤ect, or a
testing the link between task di‰culty and cerebral change in strategy. As we pursue perceptual learning
blood flow, Grady et al. (1996) found a negative studies, we must make sure that what is learned is
correlation between accuracy and rCBF in striate under strict control—that learning represents a true
cortex. They suggested that reduction in that area sensorial improvement.
might improve performance on the task (a match-to-
sample task of progressively degraded faces). Using
fMRI, Vaina et al. (1998) found that brain activa-
tion specific to the discrimination of the direction of
motion decreased in various areas after subjects had
improved in this task.
Other research groups reported increases in the
activation after learning a task. Karni et al. (1995)
observed an expansion of the activation in primary
visual cortex measured by fMRI, after training hu-
man subjects in a task to identify the orientation
(horizontal or vertical) of a small target block (an
array of three diagonal line elements di¤ering only
in their orientation from a background of identical
elements) (cf. chapters 4, 6). They strongly suspected
that the neuronal correlates for this type of percep-
tual learning would be found in primary visual cor-
tex because of the specificity of the learning e¤ect:
This page intentionally left blank
Perceptual Learning and the Development of
Complex Visual Representations in Temporal
Cortical Neurons
6
David L. Sheinberg and Nikos K. Logothetis
Abstract For example, the first time one visits a new city,
perhaps viewing its sights from the back of a taxi, the
At present, our understanding of the static properties of scenery is novel, the buildings unfamiliar, and the
temporal cortical neurons, though more advanced than our streets confusing. However, after a number of visits
knowledge of how the properties of these cells change
things begin to look di¤erent. The initial impressions
over time, is still sorely lacking. In this chapter, we there-
fore review data from physiological studies that give some fade, and are instead replaced by concrete views of a
indication of what these cells might be doing, or at least place well known. At its peak, familiarity may even
what they might represent. If learning indeed modifies the cause things to go unnoticed. Amazingly, this trans-
properties of these neurons, then the snapshots of a cell’s formation from a primarily sensory experience to
life provided by these standard single-unit studies can tell
a recognized encounter usually happens e¤ortlessly
us something about what is learned. We present evidence
that perceptual learning may provide a useful framework and without conscious intervention. Merely seeing
for understanding inferotemporal (IT) neurons. We argue something over and over can powerfully change how
that when novel patterns of su‰cient complexity are we see that same thing. In this view, every visual
regularly encountered, long-term changes occur in the encounter has two related, but independent, con-
connectivity of IT cells that implicitly incorporate these sequences: one perceptual and one mnemonic. In
experiences. We do not suggest that the activity of these
the brain the same cells whose activity causes a per-
specialized cells ever replaces the activity of cells in earlier
visual areas, but rather that it o¤ers compact and reliable ceptual experience are also possible participants in
representations that can be used for rapid perceptual anal- perceptual learning.
ysis and that can be associated with both actions and other The problem of perceptual learning is not new.
neuronal representations. More important, we think that Philosophical interest in such learning dates back to
this experience-dependent modification is not a special
at least Locke’s empiricist theories, and psychophy-
process that is explicitly turned on and o¤, but is instead
always operational during active (and perhaps even passive) sicists have been studying such phenomena system-
visual processing. Thus object learning and object recog- atically for over a century. In the field of visual
nition may not be so easily distinguished. sensation, for example, Volkmann (1858) inves-
tigated how subjects’ thresholds for two-point dis-
crimination changed with experience. He found that
a few hours of practice decreased the minimum de-
6.1 Introduction
tectable separation between two points of stimu-
When we are faced with something visible for the lation in these subjects by nearly 50%, and that
continued practice, over days and weeks, further
first time, recognition is not a prerequisite for per-
ception or awareness. But if the object is encoun- improved their performance. In light of such data,
tered regularly, its appearance does seem to change. Gibson and Gibson (1955) asked what happens in
the course of this type of learning. They forcefully
96 David L. Sheinberg and Nikos K. Logothetis
resisted the notion that experience causes perception We, however, find the processes underlying vision
to become less a function of the stimulus and more a far more entangled than this. Where perception ends
result of the observer’s enrichment of that stimulus and recognition begins is almost certainly not so
by internal processes. Instead, they proposed that the neatly defined. Traditional tasks in visual perception
intrinsic properties of a well-learned stimulus be- (orientation discrimination, contrast detection) gen-
come more closely associated with the percept, and erally require that the subject map a visual experience
that performance on discrimination tasks improves as to a response, and this mapping requires that some
a result of more specific identifying responses. ‘‘We internal representation of the stimulus be connected
suggest that the stimulation is complex, not simple, to explicit reactions ( judgments, responses). On the
and that the observer continues to discover higher- other hand, classic paradigms in the domain of visual
order variables of stimulation in it. The percept recognition, such as visual priming, are primarily
becomes di¤erentiated’’ (Gibson and Gibson 1955, aimed at probing the brain’s internal representation
40). According to the Gibsons, there is no need to of specific visual stimuli, and in these studies, the
add information to an already rich stimulus. Instead, actual response is of only secondary interest. The
the brain becomes more capable of extracting the addition of learning or memory into the equation
information that was there all along. further complicates the matter with regard to a sim-
In a book on perceptual learning, it is reasonable ple parceling of visual function. Task demands and
to ask whether the problem of visual recognition experience seem to play an important, and often
qualifies as a form of perceptual learning at all. Basic overlooked, role in not only classification and cate-
treatments of vision often divide the problem into gorization, but in perceptual analysis as well. One
three distinct stages: sensation, perception, and rec- issue that remains paramount, though, is the ques-
ognition. In this view, sensation comprises the initial tion of representation. How does the brain represent
stage of sensory transduction, carried out by photo- a visual stimulus? In considering this, we can also ask
receptors in the retina. The raw sensory signal, which how this representation changes with experience.
we know is transmitted in large part to visual cortex From this perspective, we feel that a chapter devoted
through the thalamus, is then transformed into dis- to learning of complex visual patterns and the neu-
tinct percepts, probably as a result of complicated ronal correlates of this process does indeed belong in
interactions between multiple visual areas. Finally, a book dedicated to the study of perceptual learning.
the completed percept may trigger a reaction (either This view is not new. In arguing that perception
physical or mental). In such a simplified view, learn- is governed by the same processes as more concep-
ing to associate percepts with reactions (recogni- tual activities like categorization, Bruner (1957, 124)
tion) might well be considered a process separate wrote: ‘‘A theory of perception, we assert, needs a
from learning to perceive. A similar view is found mechanism capable of inference and categorizing as
in theories of computational vision, which often much as one is needed in a theory of cognition. . . .
divide the problem of vision into early-level, middle- [I]t seems to me foolish and unnecessary to assume
level, and high-level processes. Here again, percep- that the sensory ‘‘stu¤ ’’ on which higher order cate-
tion would generally map onto middle-level vision, gorizations are based is, if you will, of a di¤erent sen-
whereas recognition would be considered a high- sory order than more evolved identities with which
level process. our perceptual world is normally peopled.’’ He goes
Development of Complex Visual Representations 97
on to describe a simple experiment in which two 8- one hand, neurons in this region are intimately con-
letter strings—YRULPZOC and VERNALIT—are nected with occipital brain regions that are thought
briefly flashed and subjects are then asked to report to be essential for visual sensation. On the other
the letters that were displayed. For the first string, hand, inferotemporal cortex has major reciprocal
which poorly approximates a real English word, connections with medial temporal, limbic, and
subjects correctly reported, on average, 48% of the frontal brain areas known to be critical in mnemonic
letters. For the latter, which could be but is not an function (see Logothetis and Sheinberg 1996 for re-
English word, subjects correctly reported 93% of the view). Physiological studies have shown time and
letters. Why the di¤erence in perceptibility? Bruner again that single IT cells can be selective for complex
argues that English speakers, manifesting what has patterns, but exactly how this selectivity comes about
been described as the ‘‘word superiority e¤ect,’’ learn is still not clear, although there is increasing evidence
to code words in a way that accounts for the natural that in the mature adult, visual experience continues
variation in the language, and that under ‘‘substan- to modify the synaptic connections (Buonomano and
dard’’ conditions (e.g., tachistoscopic presentation), Merzenich 1998a; cf. chapters 2, 4). Within infero-
words that conform to this model can activate pre- temporal cortex, such modifications may create re-
existing codes and thus be perceived more veridically liable and relatively sparse codes of encountered
than words that do not (see also Reicher 1969). visual stimuli. At the same time, changes in the e‰-
Unlike the Gibsons, Bruner (1957, 127) argues that cacy of divergent connections to limbic, frontal, and
perceptual learning does indeed involve ‘‘learning’’ subcortical brain areas may provide a route for con-
on the part of the perceiver: ‘‘I would propose that necting visual encounters with appropriate reactions.
perceptual learning consists not of making finer and
finer discriminations as the Gibsons ([Gibson and
6.2 Features and Representations
Gibson] 1955) would have us believe, but that it
consists rather in the learning of appropriate modes
What do cells in the inferotemporal cortex signal?
of coding the environment in terms of its object
Such a question is obviously impossible to answer
character, connectedness, or redundancy, and then
unequivocally for this or any brain area. Even so,
in allocating stimulus inputs to appropriate categori-
for neurons in early visual areas, neurophysiologists
cal coding systems.’’
generally feel comfortable characterizing the kind of
In this chapter, we follow the direction spelled out
information transmitted by these cells using terms
by Bruner and address the question of how experi-
such as contrast, orientation, color, disparity, and motion.
ence with visual patterns might lead to new ‘‘modes
What about the activity of cells in inferotemporal
of coding the environment.’’ Specifically, we con-
cortex? The earliest published reports of the prop-
sider how these codes may be evidenced in the
erties of such cells recount problems that to this day
physiological properties of neurons. We suggest that
remain unresolved. Gross and his colleagues (Gross,
for the purposes of perceptual learning, the cortex of
Bender, and Rocha-Miranda 1969, 1305) remarked
the lateral inferotemporal (IT) lobe roughly corre-
that ‘‘by largely confining the stimuli to bars, edges,
sponding to Von Bonin and Bailey’s area TE (1947),
rectangles, and circles we may never have found
located in the middle and inferior temporal gyri, sits
the ‘best’ stimulus for each unit. There were several
in a unique position within the visual system. On
98 David L. Sheinberg and Nikos K. Logothetis
units that responded most strongly to more compli- particular, in the field of vision, these ideas seemed
cated figures.’’ What are these ‘‘complicated figures’’ to apply well to cells in the retina, thalamus, and
and how could one ever hope to identify them primary visual cortex. Hubel and Wiesel’s landmark
systematically? work (1962, 1968) in striate cortex also fit well with
the notion of a hierarchical structure providing cells
6.2.1 The Grandmother Cell with increasingly complex representations of visual
stimuli. More recent behavioral studies (see Parker
Konorski (1967) hypothesized that neurons in the and Newsome 1998 for review) have fueled interest
visual system were organized in a hierarchical fashion, in the information conveyed by single cells, sug-
with low-level, elementary feature detectors at the gesting that at least some perceptual decisions in the
base, and ‘‘gnostic units’’ at the most advanced stage. behaving animal can be predicted by the activity of
The gnostic units would be tuned to such complex single cells. But despite the intuitive appeal of en-
feature combinations that they would, through their capsulating so much information in a single neuron’s
activation, implicitly ‘‘represent unitary perceptions.’’ activity, Konorski’s gnostic units and Barlow’s car-
The most notable proponent of such a hierarchical dinal cells have met mostly with ridicule and dis-
scheme has been Barlow (1972, 1985, 1995). In the belief in the three decades since they were introduced
last three decades, his hypotheses about the relation- (see Gross 1992 for discussion). Hubel (1995, 223),
ship between single neurons and perceptual events for example, observed:
have become almost synonymous with the most ex-
Do cells continue to become more and more specialized at
treme version of this theory—the ‘‘grandmother cell’’
more and more central levels, so that at some stage we can
version—where a single cell is capable of responding expect to find cells so specialized that they respond to one
to the sight of one’s grandmother.1 It is important single person’s face—say, one’s grandmother’s? This no-
to note, however, that Barlow (1972, 390) actually tion, called the grandmother cell theory, is hard to entertain
rejects this idea along with the notion that a single seriously. Would we expect to find separate cells for
grandmother smiling, grandmother weeping, or grand-
cell controls the whole of subjective experience:
mother sewing?
The ‘‘grandmother cell’’ might respond to all views of
There is ample reason to resist the idea that a sin-
grandmother’s face, but how would that indicate that it
shares features in common with other human faces, and gle cell is responsible for representing something as
that, on a particular occasion, it occurs in a specific posi- complicated as a whole object. It is true that we still
tion surrounded by other recognizable objects? Our per- know very little about how the activity of a single
ceptions simply do not have the property of being isolated inferotemporal neuron participates in the generation
unique events as one would expect if each corresponded to
of perceptual events. In Sheinberg and Logothetis
the firing of a unique neuron.
1997, we found that there are striking correlations
Important discoveries by neurophysiologists such between the activity of single temporal neurons and
as Hartline (1940), Barlow (1950), and Ku¿er (1953) visual awareness. But we were also quick to point
clearly pointed to the fact that the activity of indi- out that while such correlations may reveal impor-
vidual neurons did convey useful information about tant information about the perceptual organization
stimuli in the external world, and the notion of a of the visual system, they do not tell us that the
neuron’s receptive field arose from such studies. In activity of any one of these cells leads to the aware-
Development of Complex Visual Representations 99
ness or recognition of individual visual stimuli (see started to appear. By far the most commonly reported
Crick and Koch 1998 for discussion on the neural cells of this type were the ‘‘face cells’’ first described
basis of awareness). Furthermore, given the large by Perrett, Rolls, and Caan (1979, 1982) and by
number of successful studies of inferotemporal neu- Bruce, Desimone, and Gross (1981). Since that time,
rons in anesthetized animals, it would be hard to cells responsive to faces have also been reported in
argue that activity of these cells alone caused aware- areas outside inferotemporal cortex, such as in the
ness of anything. Indeed, from this perspective, the amygdala (Rolls 1984) and the frontal lobes (O Sca-
grandmother cell theory appears wrong. Even if laidhe, Wilson, and Goldman-Rakic 1997; Pigarev,
there were cells that responded to specific objects, Rizzolatti, and Scandolara 1979). Because the gen-
the activity of these cells may have little or nothing eral properties of these cells have been reviewed in
to do with the experience of ‘‘seeing’’ that object. detail elsewhere (Desimone 1991; Gross 1992; Log-
It has been shown, for example, that in the human othetis and Sheinberg 1996; Perrett, Mitslin, and
medial temporal lobe there are cells selective to Chitty 1987; Rolls 1994), our principal aim here is
responses to faces and expressions, and that some to address a subset of the outstanding issues that re-
such cells respond di¤erently to previously presented late to the question of perceptual learning.
stimuli, even when the subjects denied having seen An important point to note is that the category of
the stimulus before (Fried, MacDonald, and Wilson ‘‘face cells’’ is actually quite diverse and almost cer-
1997). tainly comprises multiple subclasses (Perrett, Mitslin,
and Chitlin 1987; Tovee 1995). The diversity of
6.2.2 Complexity, Selectivity, Generality, and responses can, at least in part, be traced to the di-
Reduction versity of visual areas within which face cells have
been found (Baylis, Rolls, and Leonard 1987). These
At the time Gross and his colleagues began inves- visual areas include the superior temporal polysensory
tigating the properties of visual cells in the temporal (STP) area, found within the upper bank and fundus
lobe, they were ‘‘prepared to find IT cells that fired of the superior temporal sulcus (STS), and infero-
selectively to complex stimuli such as hands and temporal cortex, in the lower bank and lip of the
faces’’ (Gross 1994, 465): they were fully aware of STS and the middle and inferior temporal gyri (area
Konorski’s hypothesis (1967) that inferotemporal TE).
cortex may be site of the controversial ‘‘gnostic Specific testing of face selective cells has shown
units.’’ The first published study on the physiological that the responses of some of these neurons are de-
properties of IT cells (Gross, Bender, and Rocha- pendent on head orientation but less so than they are
Miranda 1969), however, measured the response of on the identity of the face (Desimone et al. 1984;
IT neurons to relatively simple visual stimuli, such as Oram and Perret 1992; Perrett et al. 1985). Like-
di¤use light, bars, and edges. Only reluctantly, Gross wise, other cells seem tuned to particular facial
(1994) recounts, did they include anecdotal infor- expressions (Hasselmo, Rolls, and Baylis 1989). Fig-
mation about a cell that seemed to respond best to a ure 6.1 shows an example of a cell of the former
monkey hand. Almost an entire decade passed after kind, and it also illustrates a number of properties of
the initial IT physiology studies before more reports many of the temporal cortical neurons we—and
of cells responsive to complex biological objects many others—have encountered (see figure caption
100 David L. Sheinberg and Nikos K. Logothetis
Figure 6.1
Response profile of a typical ‘‘face cell,’’ which responded most vigorously to profiles, but was sensitive also to the subject in
the profile. Note the large dynamic range in firing frequency between responses to preferred and nonpreferred (e.g., the
frontal face views) stimuli. For all cells shown in this and the following figures (except figure 6.6), each neuron was tested
with a series of briefly presented visual stimuli. The data in these figures were collected from two rhesus macaques, using
surgical methods and chamber placement described in Sheinberg and Logothetis 1997. The activity of each cell was recorded
while the monkey maintained fixation within a region approximately 2 square. Stimuli were approximately 4 on a side.
None of the stimuli in this or the following figures (except for the wire objects in figure 6.6) was ever designated as special to
the monkeys. Each graph represents the response of the neuron to the image shown beneath. Vertical ticks at the top of each
plot denote a single, well-isolated, spike; the vertical lines show stimulus onset and o¤set times. The perfectly aligned hor-
izontal ticks to the left of the stimulus onset line indicate a row corresponding to a single trial (useful for trials that contain no
spikes). The firing rate for each stimulus condition is estimated as in Sheinberg and Logothetis 1997, and shown as the filled
curve in each graph.
Development of Complex Visual Representations 101
for a description of the basic methods).2 First, the cell the left profile, and a substantial response to the right
exhibits an enormous dynamic range of firing fre- profile. The response to the two views is not iden-
quencies. Its spontaneous rate is quite low, firing, on tical, however. This tuning demonstrates that while
average, below 1 Hz prior to the presentation of a cells are ‘‘selective’’ with respect to the preferred set
visual stimulus. Although this rate leaves little room of stimuli, they also exhibit some generality for other
for inhibition, the cell’s response to frontal views of dimensions. The details of other forms of general-
either humans or monkeys appears to fully inhibit ization in IT cells has been investigated in numerous
the cell: no spikes occurred from 120 msec after studies by systematically varying basic image attrib-
stimulus onset (the cell’s latency) until stimulus o¤- utes of the test stimuli, including both face and non-
set, when these views were presented. The cell’s most face objects (e.g., Fourier descriptors and wirelike
vigorous response was in excess of 100 Hz, achieved objects). These variations include changes in contrast
shortly after its onset latency. Note also that after ef- (Rolls and Baylis 1986) and position (Ito et al. 1995;
fective stimuli, cells often continue to respond, even Logothetis, Pauls, and Poggio 1995; Schwartz et al.
though the stimulus is no longer physically present. 1983; Tovee et al. 1994), where generalization is
Second, the cell’s response modulation is quite quite robust, as well as in scale (Ashbridge et al.
consistent from trial to trial for at least some of the 2000; Ito et al. 1995; Logothetis, Pauls, and Poggio
stimuli tested, although this reliability is limited to 1995; Schwartz, et al. 1983), spatial frequency con-
short epochs starting soon after the cell’s overall on- tent (Rolls, Baylis, and Leonard 1985), color (Perrett,
set latency. Tovee et al. (1993) have shown that a Rolls, and Caan 1982), and lighting (Hietanen et al.
substantial portion of the information transmitted by 1992), variations that can lead to more dramatic
IT cells can be extracted from intervals as short as 50 changes in cell response.
msec. Which parts of the spike train are most critical Although the variation with head orientation
is clearly an important, but unresolved, issue. Rich- seems to capture the essence of what the cell in figure
mond and his colleagues (Richmond and Optican 6.1 signals, dubbing it a ‘‘profile cell’’ does not seem
1987; Richmond et al. 1987) have argued that the entirely justified by its pattern of response. It cer-
temporal characteristics of the response must be taken tainly can contribute to a process of discriminating
into account to fully describe the information con- head orientation. But the response graphs and pro-
veyed by single units (cf. chapter 5). We, however, files shown in the bottom two rows show that the
would suggest that because most recognition prob- cell does distinguish between the various heads in
lems can be solved on the order of hundreds of milli- profile. It may come as little surprise, then, that the
seconds, and because unconstrained fixation durations most e¤ective stimulus we found for this neuron was
last, on average, about 250 msec, the critical infor- a profile of an individual who was well known to
mation conveyed by IT neurons for the purposes of the animal (the scientist depicted in the top row of
perception and recognition must be conveyed as figure 6.1). It is commonly assumed that if and when
early and as rapidly as possible (see also Oram and high selectivity exists, it should be reserved for objects
Perrett 1992). of extreme importance to the animal (such as the
Finally, for the cell in figure 6.1, there is a sys- experimenter charged with its well being). Although
tematic relationship between head rotation and the such a view seems reasonable, we think it is probably
cell’s activity, with the maximal response elicited by not entirely correct. Figure 6.2 shows a cell that
102
Figure 6.2
Highly selective inferotemporal (IT) neuron, which responded best to a single image from our test set. The contrast reversed image had
never been seen before the trials depicted here. Note that the spontaneous activity of this cell, like many others we have encountered in IT
cortex, was very low. There was generally no discernable response to nonpreferred faces. (See caption for figure 6.1.)
David L. Sheinberg and Nikos K. Logothetis
Development of Complex Visual Representations 103
Figure 6.3
Di‰culties with the reduction method. (a) This cell responded vigorously and reliably to a particular image of a monkey
sitting on a rock, as well as its mirror image (both of which had been seen before this recording session). Response to other
views of the same monkey elicited almost no detectable response. (b) Cropped images showing that individual parts—head
and tail (at the same size as the original whole body image)—evoked a partial, but significant response from the cell. (c)
Original image from which the tail in panel b was extracted had little excitatory e¤ect on the cell’s response, which shows
that the essential features responsible for activating a cell cannot be reliably determined without placing those features in a
di¤erent context. Here we would argue that the cell responded to the subparts shown in panel b as a result of its learned
response to the images in panel a, and not that its response to the whole figure was a lucky consequence of various subparts
appearing together. (See caption for figure 6.1.)
Development of Complex Visual Representations 105
In any case, we believe that the notion of ‘‘critical tivity for the specific objects that had been learned
features’’ found using the reduction technique is in- during training. More recently, we have found that
appropriate for a cell that has developed its selectiv- even in the absence of extensive training, unusually
ity for a particular configural stimulus. high selectivity can still be found for images that have
merely been observed in the past (see also Booth and
6.2.3 Biological Relevance Rolls 1998). It thus seems that extensive experience
with configurally complex objects may alone account
Although configural e¤ects were originally thought for the special processing and special cells originally
to be unique to face processing, similar e¤ects have ascribed to faces. Figure 6.4, for instance, illustrates
now been found in expert dog breeders when rec- two examples of cells selective for neither faces nor
ognizing dogs (Diamond and Carey 1986) and in hands. Instead, arbitrary images taken from the large
other experts when processing a carefully controlled set of test stimuli activated these otherwise quiescent
set of novel objects (Gauthier and Tarr 1997). In- cells. Finding appropriate stimuli for such cells re-
terestingly, our understanding of the physiological quired presenting subjects with hundreds of images
properties of IT neurons seems to be following a already shown in prior recording sessions, until at
parallel path. Once the existence of face cells was fi- least one image had some systematic e¤ect on the
nally accepted, the question arose of whether such cell’s response. Although novel images were also
cells constitute an exception in inferotemporal cor- introduced on a regular basis (after which they were
tex or whether cells with similar selectivity for other no longer novel), we never found selective responses
objects also exist there but are simply more scarce for totally unfamiliar stimuli (a point to which we
(e.g., Desimone 1991). In considering this question, will return below). On a typical penetration during
Gross (1992, 6) wrote: these explorations, anywhere from two to ten can-
didate cells had to be passed up simply because no
With the exception of face- and hand-selective cells, there
appropriate stimulus could be found.4 We will of
is no evidence for IT cells that are selective for visual
objects such as fruit, tree-branches, monkey genitalia, fea- course never know whether any appropriate stimulus
tures in the monkey’s laboratory or natural environment or actually existed for such cells.
any other object. IT neurons may well discharge to these
and other stimuli but they have not been shown to be 6.2.4 Sparse and Distributed Representations
particularly selective for them. Of course, it is possible that
no one has presented the appropriate stimulus while re-
cording from the appropriate cell. We are not suggesting that all cells in inferotemporal
cortex are as selective as the ones described above.
Subsequent studies from our laboratory (Logo- Many studies of IT neurons have concluded that
thetis, Pauls, and Poggio 1995; Logothetis and Pauls these cells, though capable of distinguishing between
1995) have shown that selective responses in infer- general classes of stimuli (e.g., faces versus nonfaces),
otemporal cortex are not limited to biological stimuli respond to many members of the preferred class
such as faces and hands. In these studies, discussed in (e.g., Baylis, Rolls, and Leonard 1985, 1987; Rolls
greater detail below, monkeys learned to recognize and Tovee 1995; Young and Yamane 1992). Such
entirely novel sets of objects. After training, a small conclusions are not incompatible with the view pre-
percentage of IT cells were found to exhibit selec-
106 David L. Sheinberg and Nikos K. Logothetis
Figure 6.4
Response to nonbiological stimuli. Highly selective responses are not limited to biological stimuli such as faces or hands
(Gross 1992). (a, b) Two neurons reliably activated by images of nonbiological objects likely to have never been seen by the
monkey before our experiments began (although, by the time these cells were found, each of these images had already been
presented during the basic fixation task many times). (See caption for figure 6.1.)
sented here, although we have chosen to emphasize images. Neither ‘‘faceness’’ nor head orientation can
the high selectivity of some cells as opposed to the explain what makes one set of three stimuli excite
mild selectivity of many others. Our purpose in do- the cell and another set not. There are no obvious
ing so is not to mislead one into thinking that in- visual features that distinguish the set of e¤ective and
ferotemporal cortex is full of neurons responding none¤ective objects. As far as useful generalization
to one and only one stimulus. Instead, we think the goes, the selectivity of this cell seems to have run
unusual selectivity of these neurons provides impor- amok. It is of course entirely possible that the same
tant clues that can guide our thinking about the cell participates in the coding of multiple unrelated
population of IT cells as a whole. In figure 6.5, we images, like bits in an ASCII code. We expect that a
present a cell that responds to the presentation of a better understanding of the processes that underlie
small subset of images, but not at all to other, similar the development of selectivity may help explain the
Development of Complex Visual Representations 107
Figure 6.5
Robust responses to seemingly unrelated images. Here we show a cell that responded well to a subset of animals, and not at
all to others. No obvious measure of similarity seems to link the e¤ective stimuli and to exclude the ine¤ective stimuli.
(Most cells did not respond to only a single object.) We see this as evidence that selectivities emerge for experienced images,
and that the images ultimately activating a cell need not be especially similar to one another. (See caption for figure 6.1.)
variety of cell properties that exist within the tem- on experience, even in the form of spontaneous
poral lobes. neural activity (see Katz and Shatz 1996 for review).
From an information-theoretic perspective, there
are important distinctions between very sparse and
highly distributed representations (see, for example, 6.3 Plasticity in Visual Cortex
Abbott, Rolls, and Tovee 1996; Barlow 1972, 1995;
The range of selectivities encountered in infero-
Rolls 1994; Rolls and Tovee 1995).5 Here our pur-
temporal cortex makes one wonder exactly what
pose is not to settle the issue of which of these codes
forces are at work in the creation of these cells. What
is most compatible with actual brain processing.
distinguishes the neural activity associated with novel
Instead, we want to pursue the idea that whatever
objects from that associated with known objects? If
representations may exist can be attributed in large
one were able to simultaneously record from all vi-
part to previous visual encounters experienced by
sual areas of the brain, which areas or populations of
the system under study. If single cells’ synaptic inter-
connections can be biased in favor of multiple stim- cells would respond di¤erently to an object once it
became familiar? It seems unlikely that the pattern of
uli, then one would expect to find a whole range of
retinal excitation elicited by a visual stimulus would
selectivities across a population of cortical neurons.
di¤er as a result of familiarity, or that such di¤er-
Many theories have been proposed regarding the de-
ences would be detectable in the responses of cells
velopment of selectivity and the stability of such
in primary visual cortex (cf. chapter 5). Indeed, it is
processes (e.g., Bienenstock, Cooper, and Munro
possible that no aspect of the visual representation
1982; Kohonen 1982; Linsker 1988), some of which
of a stimulus would change as a result of stimulus
have received considerable neurophysiological sup-
familiarity. In such a scheme, memory and recogni-
port (Kirkwood, Rioult, and Bear 1996). One thing
tion would be completely separated from perceptual
these theories have in common is their dependence
108 David L. Sheinberg and Nikos K. Logothetis
analysis. The latter could depend on a set of immu- Schyns and Rodet 1997) have recently been able
table, universal primitives that would be rich enough to demonstrate that categorical judgments based on
to represent all visual objects: although the particular perceptual features critically depend on task demands
combination of active cells would be novel for a and subjects’ experience with the stimuli. For ex-
stimulus encountered for the first time, subsequent ample, Schyns and Rodet (1997) devised a set of vi-
encounters with the same stimulus would lead to sual stimuli they called ‘‘Martian cells,’’ which were
nearly identical activation patterns. Just as com- circular backgrounds textured by amorphous blobs.
binations of letters of an alphabet can represent all Categories were determined by the presence of cer-
words in a language, these visual primitives would tain characteristic blobs. Two separate blobs (x and
be capable of representing any object, regardless of y), as well as a conjunction of the two (xy) were the
whether the object was novel or familiar. defining features of three separate categories of cells
(X, Y, and XY ). The experimental question was
6.3.1 Development of Representations whether learning to visually categorize XY cells be-
fore learning to categorize X or Y cells would e¤ec-
But what if the visual brain were less static and in- tively establish the combination xy as a ‘‘primitive’’
stead continued to create new visual primitives for feature. The authors tested this hypothesis using
representing visual stimuli? Such a process would be objects that contained the individual features x and y
analogous to creating representations for words that presented as parts of the same object, but not con-
were as ‘‘primitive’’ as representations of the letters joined (X-Y cells). Subjects who had learned the
that compose them. The ‘‘word superiority e¤ect,’’ conjunction first (XY_X_Y ) categorized the X-Y
briefly alluded to in section 6.1 (see, for example, probe stimuli as either being an X or a Y, indicating
Reicher 1969), in which letters imbedded in words that the feature xy was part of these subjects’ feature
are better recognized than letters in isolation or letters vocabulary as a single unit that could not be broken
in nonwords, provides behavioral support for this apart. In contrast, subjects who had learned the in-
notion. This kind of e¤ect indicates that experience dividual features before learning the conjunction
may play an important role in reorganizing percep- (X_Y_XY ) categorized the X-Y probes as XY cells,
tual processes. It is perhaps surprising, then, that nu- suggesting the XY class was based on a second-order
merous influential models of visual recognition are conjunction of the primitive x and y features. These
based on the idea of structurally constraining the data imply that categorical judgements can be sys-
space of all objects using a fixed set of primitives tematically influenced by experience, and that per-
(e.g., Biederman 1987; Marr and Nishihara 1978). ceptual analysis is clearly subject to learning e¤ects
Recent treatments of the basic units of perceptual (Schyns, Goldstone, and Thibault 1998).
processing are clearly reconsidering the notion that The benefits of a flexible and dynamic set of
primitives for vision, audition, language, and con- feature primitives are rather obvious. First, it is not
cepts are fixed or predetermined, arguing instead necessary to posit that the perceptual system is in-
that they may develop throughout life (e.g., Schyns, nately endowed with a set of primitives broad enough
Goldstone, and Thibaut 1998). to represent all possible objects. And second, if
Using carefully designed experimental stimuli, re- such a broad set of primitives did exist, it would
searchers (Gauthier and Tarr 1997; Goldstone 1998; likely be wasteful because many feature dimensions
Development of Complex Visual Representations 109
might have representations that would never be justify the third dogma by saying the evidence suggests
used (Schyns, Goldstone, and Thibault 1998). On such a dictionary may be built up, though we are far from
the other hand, it is entirely possible that the visual being able to look into its pages by physiological methods.
( p. 386)
system has inherited a set of primitives that, without
significant rewiring, may not be appropriate for solv-
6.3.2 Plasticity in Early Visual Areas
ing the kinds of problems paramount to our survival.
Because nature rarely scraps old hardware in favor of
The importance of natural experience in the devel-
a totally new design, the primitives that we often
opment of early visual areas was first studied in detail
place so much emphasis on (such as oriented line
in the cat by Wiesel and Hubel (1963, 1965) and in
segments) may need to be modified before they can
the monkey by Hubel, Wiesel, and LeVay (1977).
e¤ectively solve more sophisticated visual needs. As
Demonstrating the importance of normal visual ex-
an analogy, consider analyzing the individual pieces
perience in the development of binocular repre-
that make up a World War II aircraft. The compo-
sentations in striate cortex, these visual deprivation
nents that went into building these planes may have
experiments served as a springboard for literally
been manufactured in a converted factory that orig-
hundreds of studies of cortical plasticity in both in-
inally built washing machines rather than flying
fant and adult animals, which have repeatedly shown
machines. The parts may thus have properties that
how damage or surgical manipulation of the normal
were optimized for a totally di¤erent class of prod-
inputs to primary somatosensory, auditory, and visual
ucts, but, when assembled appropriately, perhaps
cortices can lead to significant synaptic remodeling
with ad hoc modifications, they may have been ade-
(see Buonomano and Merzenich 1998a, chapter 2).
quate for constructing the wings, bodies, and engines
In visual cortex, such reorganization is most easily
needed to build reliable airplanes.
induced in early ‘‘sensitive periods’’ (Hubel, Wiesel,
Is such a dynamic, feature creation scheme con-
and LeVay 1977; Wiesel and Hubel 1965), during
sistent with the known physiological properties of
which both topographic representations and basic
the visual system? The ‘‘third dogma’’ of Barlow’s
neurophysiological properties, such as eye dominance
neuron doctrine for perception (1972) addresses pre-
and orientation preference, can be altered (see In-
cisely this point. In it, he proposes that selectivity of
troduction). That these changes are more di‰cult to
individual cells adapts to the environment, both
find in adult animals is not surprising if one accepts
through genetic predispositions as well as through
the tenet that the basic flow of visual information
‘‘plasticity of the neural structures involved’’ ( p. 385).
primarily feeds forward out of area V1 and into
Referring to neuronal representations as words in a
extrastriate areas. In such a hierarchical system, it is
dictionary, Barlow asks:
critical that reliable coding schemes at lower levels
Are the dictionary words there, only the ones experienced remain firmly in place. Otherwise, operations de-
becoming permanently connected; or do the cells them- pendent on information derived in these areas would
selves determine that a frequently experienced pattern, no longer make sense. (cf., however, chapter 20 for a
such as lines of a particular range of orientations, are events
possible solution).
for which words are desirable? The evidence favours mod-
ification, and the idea to which it leads of the successive Probably the most influential theory regarding the
hierarchical construction of a dictionary of meaningful mechanism of synaptic modification is Hebb’s pos-
neurons has enormous appeal. For the present we can only tulate (1949), in which he states that that coactivation
110 David L. Sheinberg and Nikos K. Logothetis
of an excitatory input and its postsynaptic target will Kato (1997) compared activity-dependent plasticity
lead to increased e‰cacy of that input. It should come in area V1 and inferotemporal cortex in an anes-
as no surprise to many, though, that William James thetized adult monkey (Macaca fuscata). Interestingly,
(1890/1950, 567) in his monumental Principles of Psy- they found that an extracellular tetanic stimulus ap-
chology proposed a similar theory in what he called plied every 4 sec for 3–5 min had opposite e¤ects
the ‘‘elementary law of association’’: ‘‘The amount of in areas V1 and TEd (dorsolateral area TE). Stimula-
activity at any given point in the brain-cortex is the sum of tion in the supragranular layers of area V1 led to a
the tendencies of all other points to discharge into it, such decrease in the evoked extracellular field potential
tendencies being proportionate (1) to the number of times the (EFP) that lasted at least 3 hours. In TEd, however,
excitement of each other point may have accompanied that of similar stimulation induced a gradual potentiation
the point in question; (2) to the intensity of such excite- in the EFP’s slope and amplitude that lasted for the
ments; and (3) to the absence of any rival point function- duration of the recording session (up to 4 hours).
ally disconnected with the first point, into which discharges Neurophysiological di¤erences in synaptic modifi-
might be diverted’’ (emphasis in original). ability between striate and temporal cortical areas are
Hebb’s postulate was framed using more modern consistent with a reported neurochemical gradient
anatomical references (axons, somas, dendrites, and in the same visual areas. Nelson et al. (1987) found a
synaptic knobs) and was su‰ciently explicit to systematic increase in the phosphorylation of two
warrant naming synapses that exhibit modifiable protein kinase C (PKC) substrates between striate
connection strengths as ‘‘Hebbian.’’ Experimental and temporal cortical visual areas. Phosphorylation
support for the existence of such synapses has come of one of the PKC substrates, the homologue of the
predominantly from slice studies in hippocampus, phosphoprotein F1 in the rat, is known to be critical
where, in a typical experiment, a brief conditioning in the induction of long-term potentiation (Rober-
stimulus delivered to a specific set of presynaptic son, English, and Sweatt 1996). Thus at least one
fibers results in a specific strengthening of their post- line of neurochemical evidence suggests that there is
synaptic connections (see Brown, Kairiss, and Keenan greater opportunity for plasticity, possibly through
1990; Katz and Shatz 1996 for reviews). It has been the growth of presynaptic terminals, at higher stages
shown that this e¤ect, called ‘‘long-term potentia- in the ventral visual pathway (Nelson et al. 1987).
tion’’ (LTP) is both input specific and long lasting, Although ablation studies in 1950s and 1960s
and its molecular mechanisms have now been in- established that inferior temporal cortex is essential
tensively studied (Roberson, English, and Sweatt for normal visual discrimination learning (see Dean
1996). LTP is not limited to the hippocampus, 1976 for review), only recently have neurophysio-
though, and has been recently demonstrated in visual logical studies been able to demonstrate that changes
cortex (Kirkwood and Bear 1994; Otsu, Kimura, and associated with visual learning can also be found in
Tsumoto 1995, chapters 3, 7, this volume). the activity patterns of single IT neurons. Rolls and
his colleagues (1989) were the first to look for such
6.3.3 Plasticity in Higher Visual Areas changes in the activity of face-selective cells as a re-
sult of experience. They found that the response
In one of the few direct studies of synaptic plasticity patterns for 6 out of 22 neurons changed in the short
in extrastriate visual areas, Murayama, Fujita, and period during which novel faces became familiar. In
Development of Complex Visual Representations 111
a second experiment, they found that 5 out of 26 subset of these views. The figure also demonstrates
neurons changed their response to familiar faces fol- how selective these cells can be: none of the other
lowing the presentation of a single novel face. The wire objects, or other images, activated the cell sig-
results show that face-selective neurons are not static nificantly above its baseline rate. These results have
filters, and instead seem capable of adjusting their important implications for models of object recog-
response selectivities as a result of experience. Be- nition because they provide neurophysiological evi-
cause the e¤ects reported were in the relative re- dence in support of view-based representations for
sponse of the neurons to the set of face stimuli, these visual objects (Logothetis and Pauls 1995; Logothetis,
changes seem to reflect a reorganization of ensemble Pauls, and Poggio 1995; Perrett et al. 1984; Poggio
codes in inferotemporal cortex, and not the creation 1990). In addition, they provide some of the strongest
of totally novel representations. Faces as objects were evidence to date for the modifiability of cell selec-
presumably well represented before the experiments tivity as a function of experience.
began, and thus the incorporation of new exemplars Further support for the idea that cells become
into existing visual representations makes good the- selectively tuned for objects through learning comes
oretic sense. from a study by Miyashita, Date, and Okuno (1993),
But what happens when one is required to make who trained three adult monkeys in a matching task
visual judgments about a class of objects that has with a set of 97 fractal patterns, and then recorded
never been seen before? Logothetis and Pauls (1995) the response of IT neurons to both the old set and a
addressed exactly this issue by training monkeys to novel set of fractal stimuli. In 14 out of 15 cells fully
recognize wirelike and spheroidal objects, similar to tested, the maximum response of the cells was to
objects used in previous psychophysical experiments learned stimuli, a result that could be attributed
in humans (Bültho¤ and Edelman 1992; Rock and neither to chance nor to the nature of the stimuli
DiVita 1987). The monkeys’ behavioral performance because they were generated using exactly the same
on these tasks was remarkably consistent with human algorithm. In a subsequent study, Sakai and Miya-
results; it showed that initially recognition was view- shita (1994), having trained monkeys to recognize
point dependent, and that only after training could computer-generated Fourier descriptors in a paired-
objects be recognized from all directions. The most associate task, then tested neurons that exhibited any
striking result of these experiments, though, was that pattern-selective response to their stimuli, using the
after training, individual neurons located in anterior original stimuli and parametric transformations in
inferotemporal cortex, near the anterior medial tem- the test patterns. They reasoned that if the activity of
poral sulcus (AMTS), were found to exhibit selective the cells they were recording from had been shaped
responses to test objects that the monkeys had pre- by training with the specific patterns in the learning
viously learned. Figure 6.6 illustrates the responses of set, then alterations in the test patterns should lead to
one such neuron, selective for a wirelike object. In a decrease in cell response: the transformed stimulus
the two top rows, the tuning of the cell is depicted, would no longer match the learned pattern opti-
which shows a remarkable similarity to the view mally. If, on the other hand, the pattern selective
tuning of the face cell shown in figure 6.1. Note responses were simply a manifestation of selectivity
that, although the monkey was capable of recogniz- that was always present in the connectivity of net-
ing all views of this wire, the cell responded to only a work of cells under study, then there would be no
112 David L. Sheinberg and Nikos K. Logothetis
Figure 6.6
Example cell, taken from Logothetis, Pauls, and Poggio 1995, responsive to a particular view of a novel object. This cell
provides convincing evidence that cell responses in inferotemporal (IT) cortex may be altered with experience. In contrast to
the stimuli shown in the other figures, the top two rows of wire objects depicted here were behaviorally relevant to the
monkey from whom the cell was recorded. He had learned to discriminate highly similar wirelike objects from each other,
which may explain why almost 9% of the visually responsive cells from anterior IT cortex in this monkey responded to
Development of Complex Visual Representations 113
reason to expect that the learned pattern should be eral, for the 28 out of 131 cells that responded best
a better stimulus than nearby transforms. The results to the training stimuli, the response was not sharply
clearly favored a model of selective tuning of cells: tuned to only one of the training stimuli. On aver-
the altered forms always elicited a weaker response age, 3 out of the 28 training stimuli elicited at least
than the learned image. These data are also totally a half-maximal response. This result, which can also
consistent with the findings of Logothetis and Pauls be seen in the data shown in figure 6.5, strongly
(1995), who reported that no selective cell was ever suggests that learning e¤ects do not require that cells
found to prefer an unfamiliar test stimulus. become selective for individual images. The details
Figure 6.7 depicts an example of our testing to of selective tuning are not well understood, and many
show how a cell responds to digitally altered variants factors are likely to play a role in determining the
of its most e¤ective stimulus. The top row shows the distribution of cell selectivity. An informative aspect
original image, along with five variants. Notice that of the Kobotake, Wang, and Tanaka (1998) result is
no variant proved more e¤ective than the original, that the cells were recorded in anesthetized mon-
although it is clear that the cell does continue to re- keys. The di¤erences between the trained and con-
spond quite vigorously to a number of the altered trol monkeys show that experience can lead to
images. Any time we have tried such a test, we, like persistent biases within inferotemporal cortex that
Sakai and Miyashita (1994), have never been able to are not dependent on arousal, attention, or aware-
create on the fly a more e¤ective stimulus than ness. On the other hand, the conditions for inducing
one that had been encountered before. If cells in in- such changes may well depend on all three.
ferotemporal cortex were elements of an elaborate, What conditions promote experience-based mod-
but fixed, coding system, then there would be no rea- ification in IT cells? Discussions about the potential
son to expect that our randomly chosen test images importance of past experience and training in deter-
would be more e¤ectively represented than unen- mining the ‘‘adequacy’’ of a stimulus for activating
countered images. these cells can be found in early theoretical hypoth-
In the initial studies on the coding of complex eses about IT function (Konorski 1967) and in the
objects by Tanaka and his colleagues (1991; Kobo- first neurophysiological reports of IT cells (Gross,
take and Tanaka 1994), because the monkeys were Bender, and Rocha-Miranda 1969; Gross, Rocha-
untrained and anesthetized during testing, it is not at Miranda, and Bender 1972). It has further been
all clear how much perceptual learning may have argued that the extreme importance of faces in social
taken place during these experiments. Kobotake, communication, coupled with the need to identify
Wang, and Tanaka (1998), however, have reported individual faces within the class of all faces, may be a
e¤ects of learning on cell selectivity in inferotempo- driving force behind the existence of face cells (e.g.,
ral cortex, comparing responses of IT neurons from Desimone 1991). To make fine discriminations that
control monkeys with cells recorded from monkeys cannot be based on diagnostic features, the underly-
trained to discriminate complex shapes. They found ing representations may have to be more complex in
that the proportion of cells responsive to some of the nature. One prediction is that if monkeys are taught
training stimuli was greater in the trained monkeys to make fine discriminations between objects other
than in the controls, suggesting that the characteristics than faces, then configural selectivity for these items
of the population changed with experience. In gen- may be induced. Indeed, in experiments designed
Development of Complex Visual Representations 115
Figure 6.7
Generalization for images, never before seen, that are structurally similar to an e¤ective image. The top row illustrates the
response of this cell to the most e¤ective stimulus we could find and to five altered views of that image created specifically
for this cell. The response to each of these views was significantly better than the response to most other test objects. Note,
though, that none of the variants elicited a stronger response than the original image. This finding is consistent with the idea
that the learned, e¤ective stimulus (along with other familiar stimuli, perhaps) had tuned the cell through prior experience to
respond to its specific configuration, but that this tuning was su‰ciently general so that similar images might still partially
activate the cell (Sakai and Miyashita 1994). (See caption for figure 6.1.)
116 David L. Sheinberg and Nikos K. Logothetis
to test this, Logothetis and Pauls (1995) found cells classified as face cells responded to multiple views of
responsive to images of individual objects that had one or more of the cage objects. A larger proportion
been experienced during training. Although these of the neurons responded in a view-dependent
results seem to show that very specific changes can manner to one of the familiar objects or to another
occur in inferotemporal cortex, even at the level of of the test stimuli. These results are consistent with a
single cells, they leave open the question of what model of object representations built from combi-
served as the major force underlying the change. Is nations of multiple views (e.g., Perrett et al. 1984;
it the extensive training, the extensive exposure to Poggio 1990) and are very similar to data reported
particular objects, or both that leads to the neuro- by Logothetis and Pauls (1995) for the novel objects
physiological changes? used in that study. A major di¤erence in Booth and
A study by Vogels and Orban (1994) suggests that Rolls 1998, though, is that view combination cells
extensive training on its own is not su‰cient to bias seem to have developed in the absence of extensive
the properties of individual IT neurons. They trained controlled training, suggesting that active interaction
monkeys to make fine orientation discriminations with visual objects alone may be su‰cient to alter
between successively presented gratings, and com- visual responsiveness in IT neurons. Inherent interest
pared the response of IT neurons before and after in exploring and coding the visual environment may
training. The animals’ performance clearly improved provide su‰cient exposure to particular stimuli to
with practice, but no systematic changes were ob- lead to detectable change in neural representations.
served in the neurophysiological recordings (see
chapter 5 for results in V1). An important implica- 6.3.4 Time Course of Learning
tion of this study, especially in light of the positive
e¤ects obtained by Kobotake, Wang, and Tanaka Although the existing data indicate that cell prop-
(1998), is that the visual areas involved in various erties are tuned through experience, the precise time
tasks are likely to be stimulus specific. For orienta- course for these long-term changes is not known.
tion discrimination, the cells in inferotemporal cortex When, for example, does plasticity begin? In their
may have little to contribute in such a task, and may study of macaque IT and STP neurons, cells, Rod-
therefore show little change with experience. On the man, O Scalaidhe, and Gross (1993) found that
other hand, in tasks where there may be large bene- visual cells in these areas are considerably less active,
fits to recoding complex configurations into percep- as a whole, in the infant than in the adult, and that
tual wholes, these cells may play a much more active the infant cells have longer and more variable visual
role. latencies. But they also found that even by two
One recent study has called into question the ne- months of age, single cells exhibit the same sorts of
cessity for extensive training in biasing cell activity. (varied) selectivity that adult IT cells do (see, for
In Booth and Rolls 1998, two monkeys were given a example, Rodman, O Scalaidhe, and Gross 1993,
battery of ten objects to play with in their home figure 13). Do these data imply that IT neurons se-
cages. All neurophysiological testing was done using lective for, say, faces, are genetically preprogrammed?
a fixation condition, and in no way depended on the Given that the neurophysiological properties of
monkeys’ ability to recognize the test objects. Under monkey striate cortex are quite functional at birth
these conditions, about 9% of the visual neurons not (Hubel, Wiesel, and LeVay 1977) and that mon-
Development of Complex Visual Representations 117
keys’ eyes open soon after, the opportunity for stimuli (Booth and Rolls 1998; Kobotake, Wang,
experience-based learning is clearly in place very and Tanaka 1998; Logothetis and Pauls 1995), the
early on. Further, anatomical data from Rodman and time between exposure and test ranged from several
Consuelos (1994) indicate that the pattern of visual weeks to many months. At the other end of the time
projections from earlier visual areas, including V4, spectrum, Rolls et al. (1989) reported changes in the
TEO, and posterior inferotemporal cortex, is in responses to their test stimuli in the first one or two
place in monkeys as young as seven weeks. Thus the presentations of a novel stimulus. One hypothesis
conjecture that the existence of face cells is not de- holds that the time necessary to establish new repre-
pendent on visual experience remains unsupported. sentations will be a direct function of the preexisting
Interestingly, in our adult monkey recordings, we representations at the time of exposure, coupled with
have also encountered a number of visual cells with the current task demands. We know from the exten-
long latency (greater than 200 msec), low magnitude, sive work on face cells that many cells do respond to
but selective responses—some of the same properties a wide variety of faces, and are not specialized for a
of ‘‘immature’’ neurons described by Rodman, O single face. If a novel face is encountered that can be
Scalaidhe, and Gross (1993). Three such examples adequately represented by the existing population of
are shown in figure 6.8. The responses of these cells cells (unfortunately we cannot, at present, provide a
were not tightly aligned with stimulus onset, and solid definition of ‘‘adequately’’), then there may be
both their maximum and mean firings were consid- only slight modifications of existing connectivity to
erably lower than most visually responsive cells in incorporate the new image, as reported by Rolls et
temporal cortex. Of particular interest are the data al. (1989). A direct consequence of novelty may also
illustrated in panel a: the only e¤ective stimulus we be increased visual encounters, because monkeys,
could find for this cell (the close-up image of the pill like human infants (Fantz 1964), have been shown
bottle) had first been introduced into the test set just to prefer looking at novel images compared with
five days prior. A comparison of the activity profile is familiar ones (Gunderson and Sackett 1984).
quite similar to that of a cell recorded in a 39-day- For other novel objects, the importance of iden-
old monkey (Rodman, O Scalaidhe, and Gross 1993, tifying individual exemplars may be critical. Most
figure 13b), which responded to an adult monkey monkeys, for instance, may never feel compelled to
face, but not to another infant monkey’s face or to distinguish between specific wirelike objects, just as
other control stimuli. We wonder (but will of course most people do not generally distinguish between
never know) how the response patterns of the par- specific elephants (not counting Dumbo). In these
ticular cells shown in figure 6.8 might have changed instances, specific representations for individual ex-
over the days and weeks after we recorded them. emplars may not exist. If, however, the need arises
The nature of the visual stimuli used in testing also to di¤erentiate one wire from another, or one ele-
seems to be an important factor in the development phant from another, then these representations may
of representations in these cells. Even extended ex- prove inadequate and the formation of new ones
posure to stimuli that do not contain combinations may be essential. Exactly what conditions may signal
of visual features, such as sinusoidal gratings, seems the need to create new representations is not at
to have little e¤ect on learning in these areas (Vogels all clear, though, because we have repeatedly found
and Orban 1994). In studies using more complicated selective responses for stimuli that were never
118 David L. Sheinberg and Nikos K. Logothetis
Figure 6.8
Adult inferotemporal (IT) cells with responses similar to those reported in infants. (a, b, c) The responses of the three cells
shown are strikingly similar to the kinds of responses found in infant IT cortex (Rodman, O Scalaidhe, and Gross 1993). For
each of these cells, the spontaneous firing rate was very near zero. The low overall responses and the long latency, relatively
low magnitude, and variable stimulus evoked responses are clearly di¤erent from the brisk and reliable responses seen from
many other cells. We hypothesize that these cells may have been recorded at a time when their connectivity was not firmly
set. (See caption for figure 6.1.)
Development of Complex Visual Representations 119
labeled as critical to the monkey. One possibility is ticular visual stimuli, satisfying the need for coinci-
that, though new representations at the level of sin- dent pre- and postsynaptic activation, and they also
gle cells may come and go, once such a represen- have low overall activity, perhaps making them
tation becomes critical, many others must also be prime candidates for modification.
recruited. This explanation would help explain why An exciting technique for answering some of these
so many isolated, seemingly arbitrary, selective cells questions is the use of chronic recording electrodes,
may be found alongside the more commonly en- which are designed to record from the same set of
countered face-selective cells. cells for many days or weeks. Such electrodes have
Another critical aspect to plasticity is neural already been used to record from rat hippocampal
stability. If synaptic modifications can adapt cells’ cells and somatosensory cortex (Nicolelis et al. 1989),
responses based on visual experience, what will pre- as well as from visual cortex in monkeys (Krüger
vent cells from simply staying in a state of constant 1989). Clearly, our understanding of the learning and
fluctuation? One possibility, inspired by a theoretical changes in single-cell representations may change
model, but now substantiated through neurophysio- dramatically if we are not reliant on single snapshots
logical experiments, is that thresholds for modifica- of a cell’s life.
tion may not be fixed, but may instead vary as a
function of the cells’ history (Bienenstock, Cooper, 6.3.5 Multiple Memory Systems
and Munro 1982; Clothiaux, Bear, and Cooper
1991; Kirkwood, Rioult, and Bear 1996). In this Acknowledging the growing body of evidence that
model, cells that are rarely active will be more plastic there is no universal memory system in the brain (see
than cells that are regularly active. This sliding mod- Tulving and Schacter 1990), Desimone (1992) dis-
ification threshold provides an internal source of cussed four specific ways that memory can be ex-
stability, and has interesting implications for the de- pressed within the visual system: (1) by tuning, (2) by
velopment of cells not only in primary visual cortex, association, (3) by adaptive filtering, and (4) by sus-
where it has been tested, but also in higher visual tained activity. All four ways have been found in
areas such as inferotemporal cortex. The require- inferotemporal neurons, but thus far we have con-
ments for plasticity in this theory are coincident pre- centrated mainly on one form of visual memory,
and postsynaptic activity, on the one hand, and a tuning in visual cells (see Sakai, Naya, and Miyashita
history of relatively low mean postsynaptic activity, 1994). During tuning, the basic level of modifica-
on the other. Cells are candidates for plasticity pro- tion is the representation of a single stimulus, where-
vided they have not adapted to any other stimulus as during association, multiple representations are
and can be reintegrated into the pool of candidate bound together, a process that appears critically de-
neurons if not exposed to any driving stimulus for pendent on the medial temporal lobe structures re-
a long period of time (because their history of acti- ciprocally interconnected with inferotemporal cortex
vation would be low, and thus their threshold for (Suzuki and Amaral 1994; Saleem and Tanaka 1996;
modification would be also low). It is possible, Van Hoesen and Pandya 1975). An elegant experi-
therefore, that the cells depicted in figure 6.8, for ment by Miyashita and his colleagues (1996) serves
example, may have been in the middle stages of a to illustrate the distinction between tuning and as-
plastic process. They can clearly be activated by par- sociation. In that study, monkeys were first trained
120 David L. Sheinberg and Nikos K. Logothetis
to form arbitrary associations between a fixed set of and the development of selectivity. In adaptive fil-
fractal patterns. After this initial training, the anterior tering, information about stimulus properties is com-
commissure was severed and the neurotoxin ibotenic bined with temporal information about how recently
acid was injected unilaterally into entorhinal and a particular stimulus has been seen to change the re-
perirhinal cortical areas, thus removing a large source sponse of visual neurons. Unlike tuning, adaptive
of neural backprojections into the inferotemporal filtering is almost always associated with a decrease
cortex of one hemisphere. The authors then looked in response to familiar objects (Miller, Li, and Desi-
for any neurophysiological consequences of this mone 1991; Riches, Wilson, and Brown 1991; Rolls
missing projection, using the same paired-associate et al. 1989). Early reports suggested that response
task, but now with both the old and new images. decrements only appeared when stimulus repetition
The unlesioned hemisphere served as a critical was separated by few stimuli and short time periods
control. Consistent with previous results (Miyashita (e.g., Rolls et al. 1989), although Li, Miller, and
1988), they found that in the control hemisphere, Desimone (1993) have shown that reduced responses
some cells responded to individual stimuli (both old with familiarity can bridge even 150 presentations of
and new), whereas others responded to both stimuli other stimuli, suggesting that the decrease in response
in a pair. In the lesioned hemisphere, however, the may induce attention systems to ignore recently
associated responses between paired stimuli were encountered stimuli in favor of novel visual stimuli.
eliminated, whereas selectivity for individual stimuli Moreover, the short-term e¤ects of adaptive filtering
remained intact. The results are in accord with the may be at least partly responsible for well-known
known behavioral deficits in visual associative learn- visual priming phenomena, in which the repeated
ing after bilateral lesions to the perirhinal and ento- presentation of a stimulus leads to more accurate or
rhinal cortices, which include di‰culties in forming more rapid responses to that stimulus (Schacter,
new associations and recalling or relearning old Cooper, and Delaney 1990). Kersteen-Tucker (1991)
ones (Murray, Ga¤an, and Mishkin 1993). They also found that repetition priming occurs even in the ab-
demonstrate that medial temporal lobe plays a criti- sence of preexisting memorial representations for
cal role in the expression of associations between visual shapes, although the e¤ect is limited to very
visual stimuli, and may be responsible for the view- short lags. Priming e¤ects are thus likely a combined
invariant responses to entire objects that have been consequence of both short-term adaptive mecha-
noted in inferotemporal cortex (Booth and Rolls nisms and their e¤ects on preexisting representations
1998; Logothetis, Pauls, and Poggio 1995; Perrett, that emerged through experience. This is in accord
Mistlin, and Chitlin 1987). On the other hand, se- with the basic role of codability in priming (Schacter,
lectivity for particular views of objects seems not to Cooper, and Delaney 1990; Tulving and Schacter
rely on these structures, and instead to develop from 1990), whereby experiences with objects that can be
repeated coactivation of earlier visual areas feed- e¤ectively coded by the visual system will be able to
ing into inferotemporal cortex and of the neurons facilitate subsequent encounters with that object.
within it. Sustained activity of visual cells, even in the ab-
Two other forms of visual memory, adaptive fil- sence of the physical stimulus, has often been thought
tering and delay period activity, can also be dis- to represent the short-term mnemonic process of
tinguished from mechanisms of long-term tuning keeping a stimulus ‘‘in mind’’ (Fuster and Jervey
Development of Complex Visual Representations 121
1981; Miyashita 1993; Sakai and Miyashita 1991), at the same time, generalize across them, its visual
and has generally been observed in the delay period system cannot depend on regularities that span the
between the presentation of a sample stimulus and entire space of images. Instead, visual representations
a subsequent test presentation. Recent experiments that emphasize important di¤erences of individual
have shown, however, that the sustained activity objects, while ignoring insignificant variations, are a
does not appear to depend on the need to explicitly critical component of successful visual recognition.
remember a stimulus: this activity is found not only We have argued that one method of solving this
between sample and test, but also following the test problem is to build representations of visual objects
and throughout the intertrial interval (Yakolev et al. by constantly incorporating experienced views of
1998). Indeed, for most of the cells shown in the real objects into the neural connections that together
examples above, we see that even following the make up the visual system.
stimulus o¤set, cells continue to respond if the last Specifically, that complex representations may be
presented stimulus was itself e¤ective. Thus the im- created dynamically throughout life is consistent
portance of this sustained activity is still uncertain, with the wide-ranging properties and selectivities of
although Yakolev et al. (1998) suggest that the on- inferotemporal neurons. That neurons may respond
going activity presents an opportunity for association to seemingly arbitrary stimuli, such as a particular
between neighboring views, and may have no per- view of a roller coaster, or a squirrel, or even one’s
ceptual consequence on a single trial. We know that grandmother, is not so shocking if one accepts the
the presentation of an ine¤ective masking stimulus view that response properties can be molded by
that immediately follows an e¤ective test stimulus experiences with the very same stimuli. The activity
can abruptly eliminate this sustained activity (Rolls of these neurons should not be mistaken for a rep-
and Tovee 1994). Because visually similar views of resentation of the actual object portrayed in the
an e¤ective stimulus shown in sequence do not seem picture. Instead, it would seem more judicious to
to disrupt a cell’s activity, it has been suggested that view such responses simply as a positive signal for the
associations between neighboring views may occur presence of configurations of visual patterns present
as sustained activity bridges the slight changes in an in that image. The neurons are buried in the midst
objects’ projection on the retina during visual ex- of an extensive network of cells, and only through
ploration (Földiák 1991; Wallis and Rolls 1997). interconnections with other cells can their activity
Such a model could form the basis for the develop- have any influence on either behavior or cognition.
ment of view-invariant representations. We also think that the sparseness of a cell’s response
will, in large part, be a function of its history of plas-
ticity and the organism’s prior experience with sim-
6.4 Conclusions ilar stimuli. Because neither selectivity nor sparseness
is a static property of an entire set of visual neurons,
Although it is known that natural images contain particular population tallies or general information-
statistical regularities that set them apart from random theoretic estimates may only give a small picture
noise (Olshausen and Field 1996), the set of visual about what cells in these areas are doing.
objects that can be found in these images is enormous. As we pointed out in section 6.1, Gibson and
For an organism to both distinguish among these and, Gibson (1955) argue that, in the course of perceptual
122 David L. Sheinberg and Nikos K. Logothetis
learning, observers become more responsive to the high-order representations reciprocally connecting
details of a stimulus with practice, but that they do with complex activity patterns in lower visual areas
not add something to a stimulus not actually there. can later reactivate lower-order representations, even
In no uncertain terms, they reject the idea that per- in the presence of degraded or otherwise suboptimal
ceptual learning is contingent on memory: input. Models of this sort have been extensively
studied (e.g., McClelland and Rumelhart 1981) and
There is no evidence in all of this literature on perceptual
have been shown to account well for phenomena
learning . . . to require the theory that an accurate percept is
one which is enriched by past experience, whereas a less such as the word superiority e¤ect described in sec-
accurate percept is one not enriched by past experience. tion 6.1. Indeed, many of them contain ‘‘hidden
Repetition or practice is necessary for the development of units’’ that play a role quite similar to that we are
the improved percept, but there is no proof that it incor- proposing here for cells in inferotemporal cortex.
porates memories. . . . The observer sees and hears more,
The basic benefit is thus one of predicting, filling in,
but this may be not because he imagines more, or infers
more, or assumes more, but because he discriminates more. or augmenting visual information using previous
He is more sensitive to the variables of the stimulus array. experience as guide. Because interactive processes
(Gibson and Gibson 1955, 40) guided by both data and prediction can speed rec-
ognition and thus also response, this top-down
On this issue, it is hard to conceive of a mecha-
modulation of perceptual processes may also a¤ect
nism that would change the observer’s sensitivity to
responses to the visual stimulus. Incomplete inputs
a stimulus through repetition that does not depend
can also, with the help of top-down guidance, be
on some long-term neural changes. Indeed, the data
induced to activate a response.
we have highlighted above suggest that perceivers’
A related benefit stems directly from the need to
brains do seem to reorganize as a function of visual
respond to visual inputs. Vision is not just about
experience. It is quite possible, however, that this
sensing light from the environment. Appropriate and
kind of memory is not directly related to imagina-
e‰cient reaction to this information is what makes
tion or inference, at least at a conscious level. Simi-
the process so remarkable. Although the engineer-
larly, perhaps the current use of the term memory,
ing behind high-end cameras is impressive, no one
especially with reference to its neural correlates, is far
expects a camera to do anything with the data it so
more general than the type of memory to which the
faithfully captures. On the other hand, living organ-
Gibsons refer. Their major objection was clearly to
isms must transform complex visual patterns into
the idea that perception improves by relying on in-
appropriate responses, and this mapping relies on
formation that is not directly available in the stimulus.
connections between internal representations of the
We agree that the stimulus is information rich, and
environment and the responding brain areas. Perhaps
it is precisely through interaction with information-
repeated visual encounters with a stimulus facilitate
rich stimuli that we believe perceptual learning for
not only the process of seeing a stimulus in the
recognition of complex objects occurs. One issue we
future, but also the associating of an action with that
have only barely addressed thus far is why the cre-
stimulus.
ation of these representations would be of any ben-
Direct support for this idea comes from a study by
efit to the perceiver.
Gibson and Walk (1956), who reared two sets of rats
Describing one such benefit, which he called
in controlled visual environments. The experimental
‘‘perceptual readiness,’’ Bruner (1957) argued that
Development of Complex Visual Representations 123
rats were exposed to specific visual forms (circles and basis of individual diagnostic features, then the im-
triangles) on the walls of their cages, whereas the portance of a system for configural representations is
walls of the control group’s cages were covered only minimized. Such deactivation may leave many forms
by white cardboard. After three months, both groups of perceptual processing relatively intact, perhaps so
were trained in a standard discrimination task, using intact that without careful scrutiny, these deficits
the forms present in the experimental group’s cages may go unnoticed.
as discriminanda. The results showed conclusively In humans, bilateral temporal lobe lesions have
that the rats who had been preexposed to the visual been implicated in the visual recognition disorder
forms were far more e‰cient at learning to discrim- known as ‘‘visual agnosia.’’ A number of researchers
inate the forms, as evidenced by their performance (Humphreys and Riddoch 1987; Zeki 1993) have
in a standard two-alternative forced-choice feeding postulated that visual agnosia is principally an inte-
paradigm. The experiment clearly suggests that even grative disorder. In patients with agnosia, the basic
mere exposure to sensory stimuli can a¤ect the perceptual apparatus appears intact, meaning that
process of connecting these stimuli to behavioral low-level disturbances in acuity or contrast sensitivity
responses. cannot explain the higher-level recognition deficit.
A common argument against the idea that cells Humphreys and Riddoch (1987) conclude that Lis-
may become selective for particular images in the sauer’s original distinction (1890) between appercep-
world is that there are simply not enough cells to tive agnosia and associative agnosia does not fit well
code for all the world’s objects. If we believed that in with one of their most carefully studied agnosic
order to recognize or react to visual objects one had patients, John. By Lissauer’s definition, apperceptive
to have at least one cell selective for that object, then agnosia results from problems in perceptual process-
we would be more concerned with this objection. ing, whereas associative agnosics cannot appropri-
However, our view is not that the development of ately connect perceptions to stored object memories.
selectivity in IT neurons is absolutely essential for all Humphreys and Riddoch argue that one of the
visual function. Indeed, TE lesions introduced early major tests for intact perceptual processing—a
in life lead to only mild impairment in habit forma- patient’s ability to accurately copy a visual image—is
tion and visual short-term memory tasks (Malkova, not necessarily a strong test for intact perceptual
Mishkin, and Bachevalier 1995), indicating that the processing. John, for example, can copy quite rea-
role of these cells can either be fully compensated by sonably, but he does so extremely slowly, and by
other brain areas, or that the mild impairment that paying focused attention to individual features of the
remains is a very specific form that does not disrupt source image. Similarly, when characterizing visual
all visual processing. Cooling studies (e.g., Horel et forms, John is also capable of di¤erentiating objects,
al. 1987), in which cryodes were surgically placed but he relies almost entirely on local features. He is
around the visual cortices of the temporal lobe, sup- unable to recognize familiar or famous faces, and
port the latter view. Deactivating anterior temporal even identified Winston Churchill as a woman sim-
cortical areas by cold a¤ects discriminations of com- ply because he was not wearing a tie in the picture
plex objects, such as faces, but not of simpler pat- (Humphreys and Riddoch 1987, 62). As strange as it
terns, such as oriented lines. If we follow this logic, may seem, though, Humphreys and Riddoch point
when objects can be recognized or reacted to on the out that John’s deficits are not obvious to the casual
124 David L. Sheinberg and Nikos K. Logothetis
observer because he has learned to rely on cues and 2. Although many of these observations have been made
features for solving problems that might be solved before by various investigators, because of the general
more e‰ciently by other means no longer available skepticism often associated with such data, we present a
number of examples of cells we have recorded.
to him. We think these other means may be the
formation and use of the kinds of complex repre- 3. Note that because all the test images were intermixed at
the same time, the e¤ects of short-term learning should be
sentations that exist in the anterior temporal lobe.
minimal. Indeed, for most cells, such as those presented
If we are to understand how the visual system here, we found very little persistent change in the response
e¤ectively deals with the complexity of the visual properties over the course of a 1- to 3-hour recording
world, we must begin to account for its capacity to session.
reshape itself to accommodate the details of the en- 4. Neither we nor our monkeys had the ‘‘dogged persis-
vironment and each individual’s particular needs. It tence’’ described by Hubel and Wiesel (1998, 403) that
has long been accepted that adaptability is funda- kept them going for nine hours with a single cell.
mental for survival. What is new is the growing body 5. Note that by ‘‘distributed’’ we mean the type of repre-
of data that indicates that changes in representational sentation that cannot be properly decoded without access
capacity can be traced to properties of single cells to a large number of contributing cells. Even a sparse code,
throughout the visual system. The challenge now is where only a few, relatively tuned neurons are used to
detect a stimulus, will be distributed across a brain area, but
to find ways to systematically examine these single
access to any one of these neurons will, on its own, o¤er a
neurons as they cooperate with their neighbors and reasonable description of an encountered stimulus.
as they react and adapt to stimuli that more accu-
rately reflect the richness of the real world.
Acknowledgments
Notes
synaptic interactions postulated by Donald Hebb as sensory cortex (Brodmann area 3b; Finnerty,
early as 1949. Hebb elaborated that synaptic contacts Roberts, and Connors 1999; Prescott et al. 1998)
are plastic and are modified as a consequence of and is related to learning and memory processes
simultaneous activation of the pre- and postsynaptic (Baudry 1998; Kleim et al. 1999).
neuron. More precisely, when the presynaptic action Animal studies by Merzenich and colleagues de-
potential precedes the firing of the postsynaptic monstrated as early as 1984 that the dea¤erentation
cell, the synaptic response will increase (long-term of digits (fingers) results in an altered representation of
potentiation, or LTP; see chapter 3.6.1); when the the hand in cortical area 3b. Pons et al. (1991) inves-
order is reversed, the synaptic response will decrease tigated the organization of the primary somatosensory
(long-term depression, or LTD). A typical temporal cortex in owl monkeys that had been deprived from
window for synaptic plasticity to occur varies around somatosensory input of an upper limb twelve years
100 msec. For pyramidal neurons in neocortical earlier. The authors found that the dea¤erented area,
slices, Markram et al. (1997) observed that a di¤er- which usually comprises the representation of the
ence in spike timing of 10 msec near coincidence hand, contained neurons that were responsive to the
switched the plasticity from LTP to LTD. That is, stimulation of neighboring representations, such as
every spike in a pyramidal cell could potentially those of the trunk and the face. This large-scale re-
a¤ect every excitatory synapse of a cell that was organization encompassed a cortical space across
active within 100 msec. However, there are reasons more than 10 mm in response to the long-term
to believe that synaptic plasticity is regulated in dis- dea¤erentation. In magnetoencephalographic (MEG)
tinct ways. A general form of activity-dependent studies, Elbert et al. (1994) observed the same type
regulation of synaptic transmission called ‘‘synaptic and corresponding magnitude of map alterations in
scaling,’’ recently described by Turrigiano et al. human amputees. Subsequent work revealed that the
(1998), enhances or suppresses all synaptic inputs of amount of cortical reorganization was strongly cor-
a neuron as a function of activity. The mechanism related with the amount of phantom limb pain (Flor
of multiplicative synaptic strength scaling preserves et al. 1995). Thus the loss of somatosensory input
the relative di¤erences between inputs and allows does not result in silence and degeneration of the
the neuron to adjust the total amount of synaptic respective cortical neurons but is followed by a
excitation it receives. Synaptic scaling may thus functional reorganization of the homuncular repre-
contribute to stabilization in firing rates during de- sentation of the body surface in the cortex. In addi-
velopment and in the adult brain; moreover, it may tion, increased use of a body part also alters the
help to counterbalance the destabilizing e¤ects of homunculus, leading to an expansion of the repre-
Hebbian-type synaptic modifications within the sentational cortical zones (Elbert et al. 1995) and also
neural network. According to Fregnac (1998, 845), to changes of perception such as the ability to localize
synaptic scaling can be seen ‘‘as a demonstration of di¤erent receptive fields to light tactile stimuli (Sterr
basic homeostasis, designed to return the integrative et al. 1998). These observations indicate that the
function of the cell to within a reference working capacity of the human brain for plastic reorganiza-
range.’’ Synaptic plasticity underlies alterations of tion in response to either injury or altered use of a
complex neural networks like the cortical represen- body part is retained into adulthood and is accom-
tation of the body surface in the primary somato- panied by changes in perception.
Functional Reorganization of Human Cerebral Cortex 127
ronal fields. The location of the equivalent current fMRI information to constrain source configurations
dipole (ECD) is allowed to move with time. A high to the gray matter of the cortex. This has proven to
goodness of fit of the field produced by the modeled be useful in reconstructing the cortical sheet by sur-
ECD to the real measurement provides a reasonable face rendering and in displaying anatomy, pathologi-
justification for the application of the model. More cal tissue, and functional activation in a single image.
sophisticated approaches incorporate the knowledge Biomagnetic responses to stimuli such as event-
of anatomical structures. It is known, for instance, related electric potentials (ERPs) and event-related
that regions occupied by ventricles or white matter magnetic fields (ERFs) benefit from signal averaging
lack active structures. More explicitly, for most of to enhance their signal-to-noise ratio (SNR). Data
the neuromagnetic data, only current dipoles with are generally digitized at a fixed rate to fill a data
an orientation perpendicular to the surface of gray array, while a stimulus or other synchronizing event
matter contribute to electromagnetic activity on a defines the time epoch of interest within this array.
macroscopic scale (for a review of MSI and MEG, The event is repeated and a time-locked signal av-
see Elbert 1998). erage calculated. Signal averaging improves the SNR
Equivalent current dipoles (ECD) can be localized provided the signal is invariant across trials and the
to within millimeters (Lütkenhöner 1996; Lütken- background EEG represents random noise. This is
höner, Hoke, and Pantev 1990). The accuracy of not always true (Makeig 1993). Moving visual stim-
source localization is not identical with the accuracy uli, for example, reliably produce increases in spectral
of separating di¤erent, simultaneously active sources, power in the 40 Hz band (Müller et al. 1996, 1997).
however. Whereas we can localize isolated ECDs to But because the oscillations elicited during stimula-
within a few millimeters, we can separate several tion are not phase locked to the stimulus, they are
simultaneously active sources only to within a range lost by averaging in the time domain.
about one order of magnitude smaller.
Currents flowing perpendicularly to the surface of Alterations of Somatosensory Representation
the head produce only a small signal strength in in Braille Readers
conventional MEG: they are magnetically silent, so Braille reading is an impressive example for a spe-
to speak, although they create a pronounced elec- cialized perceptual capability of the tactile sense.
trical potential on the scalp. Because magnetic fields Braille, a tactile language, is based on dot patterns
and electric potentials contain complementary infor- that can be recognized by moving the fingertips
mation with respect to their sources (Eulitz, Eulitz, slowly across the surface. For e‰cient reading, the
and Elbert 1997), the simultaneous measurement of tactile discrimination of the characters has to be fast
both signals provides additional constraints on source and correct. Experienced Braille readers can read up
localization that are not available using one type of to 200 words per minute (Foulke 1991), whereas
signal alone. Braille-naive persons find it di‰cult to distinguish
Magnetic source imaging (MSI) integrates MEG di¤erent Braille patterns and require intensive train-
and f MRI information into one data set. Whereas ing to learn to read Braille even at average level. Most
the location of a single ECD determined from MEG likely, perceptual learning is involved in this process,
data was superimposed on the corresponding f MRT and the question arises whether the improved tactual
section in earlier attempts, current techniques use skills go along with (1) an alteration of the cortical
Functional Reorganization of Human Cerebral Cortex 129
collide with an electron after traveling only one or a Neural Networks for Braille Reading and the
few millimeters. When particle and antiparticle col- Role of Primary Visual Areas in the Blind
lide, they are annihilated, and two gamma ray pho- In congenitally blind persons, the visual system can-
tons are emitted at an angle of 180 to one another not develop or adapt to visual processing. Empirical
(impulse preservation). If detectors for the gamma data suggest that the occipital cortex is active in blind
ray photons are connected in a way that only tem- individuals when somatosensory and auditory infor-
poral coincidence of the two photons is recorded, mation is processed (Kujala et al. 1997; Röder et al.
the density of the decaying substance can be recon- 1996). Sadato and coworkers (1998) employed PET
structed by means of computer tomography. As to explore the functional role of primary visual areas
noted above, 15 ObCbO is used to measure blood in congenitally blind persons and the neural network
volume. For the investigation of regional cerebral used for Braille reading. Regional cerebral blood
blood flow (rCBF), 15 O-labeled H2 O is used: the flow (rCBF) was measured in a Braille reading and
oxygen consumption can be imaged with 15 O2 . 15 O a non-Braille reading task. In the Braille reading task,
as a tracer with a short half-life of about 2 min only experienced Braille readers but no sighted con-
allows repeated measurements within one session. trols were tested, whereas the non-Braille reading
The brain mapping of a subject relies on the subtrac- task was performed by all subjects.1 Thus brain
tion of blood flow patterns between two conditions, activation to tactile tasks could be compared directly
often an active task condition and a resting condi- between groups. rCBF images were obtained by
tion. This approach has been subject to various criti- integrating the activity occurring during the 60-sec
cisms. Considering these, correlational techniques periods after tracer injection. Subtraction of task from
have been suggested as an alternative (see below). rest condition images was used to reproduce task-
Many of the techniques that have been developed related focal activity. In the blind the primary visual
for PET-based imaging have also been implemented cortex was active in both the Braille reading and the
for f MRI. To allow group comparisons, many inves- non-Braille reading task. The non-Braille reading
tigators use a normalization of the brain anatomy tasks elicited di¤erent patterns of activation in sighted
(mostly, the atlas by Talairach and Tournoux 1988). and blind subjects. In the blind, the ventral occipital
Alternatively, individual filters, averages, and scores cortex (including the primary visual cortex and the
may be extracted. One persistent problem is that fusiform gyri) was activated, whereas the secondary
global alterations in blood flow (within a subject or somatosensory area was deactivated during the non-
across a group of patients) are superimposed on task- Braille discrimination tasks. The reverse pattern was
specific changes in rCBF. Introducing global flow found in sighted subjects: the secondary somato-
into an analysis of covariance may o¤er a solution, sensory area was activated, whereas the ventral occi-
assuming that the e¤ect is linear. The assumption of pital regions were deactivated. These results suggest
linearity is not always justified, however, making that the tactile processing pathways usually linked to
interpretations sometimes di‰cult. For instance, it is the secondary somatosensory cortex are rerouted
unlikely that twice the stimulus intensity will simply in blind subjects to ventral occipital cortical regions
double the intensity of the blood flow without also that are engaged in visual shape discrimination in
changing the particular spatial distribution as a con- sighted persons.
sequence.
Functional Reorganization of Human Cerebral Cortex 131
Figure 7.4
Cortical output maps to focal right lower face muscles, right thenar muscles, and right biceps. Transcranial magnetic stim-
ulation (TMS) was applied to scalp positions arranged in a 1 1 cm grid. Grid position on the scalp can be inferred from the
location of the Cz electrode position from the international 10–20 system for EEG electrodes. Asterisks mark the positions
which, when stimulated, evoked phantom sensations in at least 60% of the trials. The schematic drawings of the lateral view
of the body depict the subject’s own drawing of the phantom limb at five weeks (wks) and four months (mo) following the
traumatic amputation. The area of the thumb ‘‘reference field’’ that evoked phantom sensations when stimulated by vigorous
rubbing is depicted in the second body scheme, four months after the amputation. (Reprinted from Pascual-Leone et al.
1996 with kind permission.)
134 Annette Sterr, Thomas Elbert, and Brigitte Rockstroh
of facial and upper extremity muscles was known. 7.3.1 Injury- and Use-Related Reorganization
Both the progression of sensorimotor reorganiza-
tion and the development of phantom sensations Damage to either the peripheral or the central ner-
were measured during an eleven-month period. The vous system can cause dea¤erentation to cortical
results are illustrated in figure 7.4. Within this neurons, which then change their original receptive
period, the organization of the sensorimotor cortex field properties and adopt the representational prop-
ipsilateral to the amputation underwent marked erties of neighboring neurons. Adjacent representa-
alterations. Before the amputation, a typical arrange- tions ‘‘invade’’ the dea¤erented field; this form of
ment of muscle representations and normal thresholds reorganization is thus also known as ‘‘injury-related
for the EMG response were found for the face, biceps, invasion’’ (cf. chapters 2, 3). Such an invasion has
and thenar muscles. Five weeks after the amputation, been demonstrated in amputees (Elbert et al. 1994)
TMS of the denervated area elicited phantom sensa- after unilateral hearing loss (Vasama et al. 1995), and
tions in the amputated limb but no contractions of spinal cord injury (Lotze et al. 1999).
face and arm muscles. Face and muscle contractions Enhanced use of a limb—and enhanced a¤erent
elicited by stimulation of the dea¤erented hemi- input condition—induces enlargement of the re-
sphere were not observed until four months after spective representation in the primary sensory cortex,
amputation. By then, stimulation of the dea¤erented a form of cortical plasticity called ‘‘use-dependent
area elicited reduced muscle contractions of biceps expansion.’’ An important prerequisite for use-
and face muscles, and the subject reported weaker dependent expansion is the behavioral relevance
phantom sensations, which might indicate the be- of the stimulation or task. Use-related expansion
ginning of an injury-induced invasion of neighbor- has been demonstrated, for instance, for players of
ing representations. After eleven months, TMS of string instruments (Elbert et al. 1995), and in Braille
the dea¤erented area revealed a total invasion by readers (Pascal-Leone et al. 1993 a, b; Sterr et al.
adjacent muscle representations. 1998).
Injury and use-related changes of cortical repre-
sentations can coexist within the same sensory sys-
7.3 Basic Aspects of Cortical Reorganization tem, as demonstrated in upper-limb amputees, in
whom the two hemispheres are subject to di¤erent
As proposed by Merzenich and Jenkins (1993, 100),
stimulation conditions. The hemisphere contralateral
cortical representations are ‘‘dynamic time-based
to the dea¤erented limb is subject to a loss of sensory
constructs,’’ formed by coincident temporal activa-
a¤erent input, whereas increased use of the remain-
tion. Animal experiments demonstrated that di¤er- ing hand increases sensory stimulation to the intact
ent input conditions lead to discernible alterations
hemisphere. These ‘‘hemisphere-specific input con-
of cortical representations. Five principles aspects of ditions’’ are mirrored in distinct alterations of the
cortical reorganization can be deduced from current dea¤erented versus the intact hemisphere (Elbert et
research: (1) injury- and use-related reorganization; al. 1997). Tactile stimulation applied to the lower lip
(2) synchronicity; (3) behavioral relevance; (4) mul- evoked responses not only in the face area, but also
tiple maps; and (5) critical periods (cf. also chapter 1). in the region that would normally respond to stim-
Functional Reorganization of Human Cerebral Cortex 135
ulation of the (amputated) hand. Ipsilateral to the totopic organization (in the example of multifinger
amputation, the hand representation was found to Braille readers presented above; Sterr et al. 1998;
increase significantly with increased use of the hand Wang et al. 1995) or to result in a fusion of repre-
as a consequence of the dependence imposed on that sentational zones (as in the case in musicians su¤ering
hand by the loss of the contralateral extremity. from focal hand dystonia; Elbert et al. 1998; Byl
Cortical reorganization following injury may personal communication).
reflect an extreme manifestation of mechanisms
involved in the use-related formation and alteration 7.3.3 Behavioral Relevance
of cortical representations (Irving and Rojan 1996).
Although this hypothesis remains to be confirmed, Use-dependent reorganization requires sensory
it is supported by evidence for the crucial influence stimulation within a behaviorally relevant setting.
of behaviorally relevant and massive sensory input Recanzone and coworkers (1992) compared the
on the self-organization of cortical representations e¤ect of enhanced input conditions on cortical
(see below; Dinse and Merzenich, chapter 2, this representations for active and passive stimulation
volume). conditions in adult owl monkeys. Tactile stimuli
(vibration with di¤erent frequencies) were applied to
7.3.2 Synchronicity a small part of the fingertip. One group of monkeys
(‘‘active’’ group A) was rewarded for learning to de-
Use-related alterations of cortical representations tect di¤erences in the frequency of a tactile vibration
depend on the temporal arrangement of the a¤erent stimulus. A second group of monkeys (‘‘passive’’
input flow. As emphasized by Clark and coworkers group B) had to accomplish an auditory task while
(1988), representations of neighboring fingers con- the tactile stimuli (identical to those in group A)
stitute ‘‘separate entities.’’ The receptive fields over- were applied. Thus, for group B monkeys, the tactile
lap along the distal-proximal axis of a single finger stimulation was not behaviorally relevant. Cortical
but not in rostrocaudal direction. This separation mapping of the somatosensory cortex (area 3b) dis-
may be due to the fact that neighboring fingers are closed an enlarged representation of the stimulated
rarely stimulated synchronously. When synchronous finger in the ‘‘active’’ group (A), whereas no ex-
stimulation is enforced, as by surgically connecting pansion of the finger representation was found in
two neighboring fingers, their distinct representa- the ‘‘passive’’ group (B). The cortical reorganization
tions can be fused. In individuals having fused fingers of string instrument players and Braille readers
(syndactyly) representations of the webbed fingers constitutes further examples of the relevance—the
share a common representation, whereas after surgi- rewarding e¤ect—of the stimulation and behavioral
cal separation of the webbed fingers, the respective response.
representations become separated (Mogilner et al. The statistics of sensory inputs (i.e., the amount of
1993). Presumably, segregated finger representations stimulation, temporal concurrence) may not be suf-
develop because individual fingers are stimulated ficient to guide changes in cortical representations
asynchronously after surgery. Synchronous temporal because the behavioral importance of the input is not
input has been found to induce a disarranged soma- necessarily related to their frequency of occurrence.
136 Annette Sterr, Thomas Elbert, and Brigitte Rockstroh
indicates that the activation of the nucleus basalis and Bloom 1997). For one example, ocular domi-
is essential for cortical reorganization (Kilgard and nance columns become elaborated by experience
Merzenich 1998). That is, use-related changes do during a critical period that coincides with rapid
not occur when activity of the nucleus basalis is overproduction of synapses in the visual cortex. For
blocked. another, the ability to readily discriminate phonemes
from languages to which an infant is not exposed is
7.3.4 Multiple Maps present only between the ages of 6 and 12 months,
but is lost later in life. Although some abilities can
Although evidence is still limited, results indicate the be regained by intensive training (Merzenich et al.
possibility that multiple maps can share the same 1996; Tallal et al. 1996), it seems unlikely that they
circumscribed cortical area in the visual and motor can be fully recovered. For example, in Elbert et al.
systems. Using f MRI, Karni et al. (1995, 1998) 1995, there was a tendency for greater cortical rep-
showed that two di¤erent keyboard tasks that re- resentation in the somatosensory cortex of players
quired the same muscle activation but di¤ered only of string instruments who had begun their musical
in their pattern of activation evoked responses in dif- training before the age of 10 years, an advantage that
ferent areas of the primary motor cortex (see also was not overridden by more intense or longer prac-
Ungerleider 1995). Evidence for multiple maps in the tice later on. Pantev et al. (1998) report similar ob-
somatosensory domain is provided by Braun et al. servations for the auditory system in humans.
1999, in which five subjects were trained to discrim-
inate tactile directional stimuli on a daily basis for four
7.4 Time Course of Cortical Reorganization
weeks. Brain mapping analysis of high-resolution
EEGs revealed that a single type of prolonged repeti-
The evidence of use-dependent cortical reorganiza-
tive simultaneous stimulation of two digits produced
tion should eventually a¤ect neurological rehabilita-
two di¤erent and opposite use-dependent e¤ects on
tion procedures. Relevant questions in this respect
the spatial relationship of the representations of the
concern the length of training and the amount of
digits in the somatosensory cortex, depending on the
stimulation necessary to induce cortical reorganiza-
nature of the discrimination condition used during
tion that results in recovery of lost abilities. Evidence
the neuroimaging sessions. The results suggest that
from animal studies suggests that two phases of re-
the two maps share the same somatosensory neural
organization can be distinguished. Short-term reor-
network specific to di¤erent modes of discrimina-
ganization is highly reversible and may occur hours
tion within the somatosensory modality. The partic-
to days after somatosensory dea¤erentation (Kelahan
ular context may determine which map dominates
and Doetsch 1984; cf. chapter 3). It is assumed that
processing.
such short-term changes of receptive field character-
istics reflect the loss of inhibition provided by the
7.3.5 Critical Periods
fibers before dea¤erentation (Alloway and Aaron 1996;
Calford and Tweedale 1988, 1990, 1991). Long-term
Many findings suggest that there are critical periods
reorganization occurs after weeks to months. In rats,
for when a certain brain organization and a certain
complete invasion of the dea¤erented area of the
ability develops (for review see Nelson 1999; Nelson
138 Annette Sterr, Thomas Elbert, and Brigitte Rockstroh
forepaw by neighboring representations was found alterations in the representation of reading finger
seven to eight months after dea¤erentation (Cusick, muscles4 after (1) one day of Braille reading under
Wall, and Wiley 1990). Comparable results have workday conditions5 and (2) after a three-day
been reported for owl monkeys (Churchill et al. ‘‘Braille-free’’ period. On both days, motor output
1998; Jain, Catania, and Kaas 1997; Kaas, Jain, and maps (to TMS) were obtained in the morning before
Florence 1998) and in a human (Pascual-Leone et al. measurement and in the evening after measurement.
1996; see above). No significant di¤erences in the sensorimotor rep-
Alterations of cortical representations are progres- resentations of the reading fingers were found before
sive in nature, so that short- and long-term repre- measurement between ‘‘workday’’ and ‘‘holiday’’
sentations most likely depict distinctive stages on that conditions. However, after one day of Braille read-
continuum. Both stages in the reorganizational pro- ing, the representation of the reading finger was
cess are presumably based on neural interactions that markedly larger than it was before measurement or
are also involved in perceptual learning (Kelahan and in the ‘‘holiday’’ condition. This demonstrates re-
Doetsch 1984; Irvine and Rajan 1996). versible use-related changes within the short time
In humans, empirical evidence for the time course period of a day. It seems interesting in this respect,
of reorganization is limited. Short-term reorganiza- that Braille readers often report that they find read-
tion within minutes to hours has been demonstrated ing much more di‰cult after ‘‘Braille-free’’ periods,
in transient dea¤erentation studies where local anes- that is, after weekends or vacations. It is possible that
thesia procedures are applied. Birbaumer et al. (1997) the cortical representation of the reading finger
employed axillary brachial plexus anesthesia in upper muscles has to be adapted over the short term, so
limb amputees to induce transient dea¤erentation that fast and e‰cient Braille reading is possible. Thus
of the stump. Cortical reorganization and phantom a ‘‘state-dependent’’ representation may exist within
sensations were recorded before, during, and after the neural network of the sensorimotor map that
the anesthesia. Within minutes, the axillary brachial can be reestablished under special circumstances. This
plexus anesthesia virtually eliminated cortical reor- ‘‘switching’’ of the map could explain why profes-
ganization as well as phantom limb pain. Rossini and sional Braille readers say they need a ‘‘warm-up
coworkers (1994) reported reversible alteration of phase’’ after a break.
finger representations following ischemic anesthesia. Similarly, training-induced alterations of cortical
This transient dea¤erentation induced an expansion representations were demonstrated in the above-
in neighboring unanesthetized finger representations mentioned study (Braun et al. 1999), which depicted
that reversed to the normal organization after the the cortical organization (1) after the first week of
e¤ects of anesthesia wore o¤. training (five training sessions); (2) after another
The phenomenon of short-term plasticity is not seven sessions (middle of the third week of training);
restricted to transient sensory input cuto¤. However, and (3) at the end of the fourth week of training. A
state-dependent alterations of cortical representa- significant enlargement of the hand representation
tions are further observed following increased use was found at the end of the training, but no di¤er-
or practice of trained skills, as demonstrated in pro- ence in size between the hand representation after
fessional blind Braille readers working full-time as the first and that in the third week, which suggests
proofreaders. Pascual-Leone et al. (1995) examined that two weeks of training are just not su‰cient to
Functional Reorganization of Human Cerebral Cortex 139
induce measurable changes in the cortical represen- remaining senses. Although data on perceptual com-
tation of the hand. In this study, the subjects trained pensation are poor, in recent years the investigation
in the task daily for one hour. In Braun et al. 1999, of the cortical aspect of cross-modal plasticity has
the subjects trained in the task for one hour; our attracted new interest.
laboratory is currently investigating whether the The most prominent example in studying cross-
same results can be obtained by a two-week training modal reorganization is blindness. The visual cortex
schedule with daily two-hour sessions. in persons blind from birth is not functionally silent,
From these results, we may infer di¤erent time but rather is employed in the processing of auditory
courses for injury-related and training-related re- and tactile tasks (Franzen et al. 1991; Rösler et al.
organization. Injury-related changes are progressive 1993; Uhl et al. 1993, 1994; Kujala et al. 1995;
in nature. After dea¤erentation, the brain continues Röder et al. 1999). If blind subjects are asked to
to adapt its cortical representations; this process perform a tactile delayed matching-to-sample task,
continues for approximately one year. Use-related enhanced EEG activity can be measured over occi-
changes can be obtained within the range of weeks, pital areas, indicating the coactivation of the visual
although early and thus more subtle alterations in cortex (Röder et al. 1996). As reported before, the
cortical reorganization may not be uncovered with visual cortex is also active during Braille reading
the available imaging procedures. (Sadato et al. 1996, 1998), and the pattern of activity
di¤ers in persons who are blind from birth and those
who became blind as adults (Büchel et al. 1998).
7.5 Experimental Evidence for Intramodal
Deafness constitutes another example of cross-
Cortical Reorganization in Humans and the modal plasticity. Auditory areas are activated by
Perceptual Consequences
tactual vibration stimuli in deaf persons, indicating
additional functionality of the deprived auditory
Functional plasticity can be divided into intramodal
cortex (Levänen, Jousmäki, and Hari 1998).
reorganization, alterations of brain functions within
In sum, the literature suggests that gross functional
one perceptual modality, and cross-modal reorganiza-
adaptations of the remaining senses can be detected
tion, alterations across di¤erent modalities. If cross-
following near-complete sensory deprivation of one
modal alterations are of interest, it is typically asked
modality. In addition, a di¤erent functional con-
to what extent the loss of one sense, say hearing or
nectivity of the deprived area itself occurs, which
vision, alters the cortical processing of the unim-
allows the activation of the deprived zones. Because
paired senses. Because it primarily a¤ects the devel-
basic functional organization of the mammalian
oping brain, we will briefly summarize findings on brain is achieved during prenatal development and
cross-modal plasticity and then examine evidence for
is very much controlled by genes, it seems unlikely
intramodal reorganization in greater detail.
that rewiring leads to a total reallocation of the
respective primary cortical areas. Nevertheless, there
7.5.1 Cross-Modal Plasticity
is no doubt that additional functional capacities are
assigned to the deprived areas, most likely via un-
According to the ‘‘compensation hypothesis’’ (Galton specific thalamocortical projections and associative
1883), the loss of a sensory modality leads to com- cortical structures.
pensatory gain in the perceptual capacities of the
140 Annette Sterr, Thomas Elbert, and Brigitte Rockstroh
sentation in area 3b. Thus use-related fusion of digit marked and significant improvement in spontaneous
representations may be involved in the etiology of repertoire performance without the splint at the end
focal dystonia. Indeed, when Elbert and coworkers of the treatment, two into the normal range. Sub-
(1998), using MSI, mapped the cortical hand repre- sequently, four subjects showed further improve-
sentation of the a¤ected hand in musicians su¤ering ment, one additional subject into the normal range.
from unilateral focal dystonia, they found it to be Three subjects retained the improvement they had
smaller than that of the corresponding una¤ected previously made. Neuroimaging results indicate nor-
hand and smaller than that in nonmusician control malization of the cortical representational maps in
subjects. This abnormal somatosensory organization successfully treated subjects. The outcome demon-
was also found in patients with writer’s cramp and in strates that learning-induced alterations in the func-
patients in whom the dystonic symptoms were asso- tional architecture of the brain can be maladaptive,
ciated with several di¤erent tasks (Bara-Jimenez et and that resulting pathology can be treated using
al. 1998). According to the Byl et al. (1997) model, behavioral techniques based on learning principles
the abated representation of the dystonic hand may derived from recent research on neuroplasticity.
indicate a fusion of dystonic fingers. Another plausi-
ble interpretation is that the perturbed motor control 7.5.3 Sensorimotor and Motor Function
in patients su¤ering from focal dystonia of the hand
is due to fused finger representations, as in the model Studies of dea¤erentation and use-related reorgani-
for multifinger Braille readers (Sterr et al. 1998). If zation in patients su¤ering from spinal cord injuries
this interpretation is valid, the segregation of the (Levy et al. 1990; Topka et al. 1991; Streletz et al.
finger representations by asynchronous stimulation 1995; Green et al. 1998; Lotze et al. 1999), from
may be a promising approach to treatment. Indeed, cortical lesions (Cohen, Bandinelli, and Sato 1991b;
Candia et al. (1999, 2000) have developed a suc- Weiller et al. 1992), and from amputations (Hall et
cessful therapy for focal dystonia of the hand using al. 1990; Cohen, Bandinelli, and Findley 1991a;
just such an approach. Their subjects were eleven Fuhr et al. 1992; Kew et al. 1994; Seitz et al. 1995;
professional musicians, six pianists, two guitarists, Ridding and Rothwell 1995; Chen et al. 1998) have
and three players on wind instruments with long- consistently found that the lesion-bound muscle rep-
standing symptoms who had previously received a resentations expand into the dea¤erented area. Tran-
variety of treatments. One or more of the una¤ected sient forearm dea¤erentation is not restricted to the
digits were immobilized with splints; the focal dys- somatosensory domain but leads to significant re-
tonic finger was required to carry out repetitive versible short-term changes of motor representations
exercises in coordination with one or more of the (Brasil-Neto et al. 1992, 1993). The amplitudes of
other digits for 112 –212 hours a day over eight con- motor evoked potentials induced by transcranial
secutive days under a therapist’s supervision. The magnetic stimulation from muscles proximal to the
subjects were instructed to continue the exercises temporarily anesthetized forearm increased within
using the splint at home for 1 hour every day for one minutes after the onset of anesthesia and returned
year after supervised treatment. The three players to preanesthesia values after the anesthesia subsided.
of wind instruments did not improve substantially, Further investigation of this phenomenon recently
whereas each of the pianists and guitarists showed revealed that plastic changes in the transient deaf-
142 Annette Sterr, Thomas Elbert, and Brigitte Rockstroh
ferented cortex can be modulated noninvasively individuals with normal hearing, which indicates
(Ziemann, Corwell, and Cohen 1998). Transcranial adaptation of the auditory system in these individuals.
magnetic stimulation of the dea¤erented hemisphere By contrast, subjective tinnitus, or ‘‘ringing in the
induces an upregulation of reorganizational pro- ears,’’ characterized by the perception of auditory
cesses. A downregulation is achieved by applying signals in the absence of any external stimulation, is
TMS to the intact hemisphere. The short-term an example of maladaptive cortical reorganization.
alterations of cortical output maps in Braille readers Cortical plasticity may account for tinnitus and
observed by Pascal-Leone et al. (1995) and described its associated symptoms (Lockwood et al. 1998).
before (see section 7.4) are an example for the cor- Mühlnickel and coworkers (1998) demonstrated the
relation of enlarged representations and highly skilled similarity between tinnitus and phantom pain phe-
perceptual capacities in the motor domain. nomena vis-à-vis brain plasticity. Their MSI-based
mapping of the tonotopic organization of the audi-
7.5.4 Auditory System tory cortex revealed an altered representation for the
tinnitus frequency, the expansion of the tonotopic
Reorganization in the auditory domain has been map being correlated with the subjective intensity of
studied less extensively than in the somatosensory the tinnitus. It is still unclear whether the observed
system. Examples demonstrating reorganization of reorganization arose from a loss of peripheral input
the primary auditory cortex include expanded rep- that may subsequently have been recovered. Never-
resentations of tones in highly skilled musicians theless, the positive correlation between the subjec-
(Pantev et al. 1998) or a structural enlargement of tive strength of the perceptual phenomenon and the
the planum temporale in musicians, as indicated by amount of cortical reorganization in tinnitus as well
f MRI (Schlaug et al. 1995). In skilled musicians, the as in phantom pain strongly emphasizes the func-
enhancement of the representations tended to be tional role of altered brain organization.
more pronounced in musicians who started musical
training early in life (Pantev et al. 1998). Use-related
7.6 Cortical and Behavioral Plasticity
reorganization of the auditory cortex was also found
Following Brain Lesions
in blind persons, who rely to a greater extent on
auditory information than sighted persons do.
Brain damage resulting from cerebrovascular acci-
Sounds that are typically neglected by sighted per-
dents can lead to complete destruction of a given
sons may be relevant for blind people, thereby im-
brain structure, combined with total loss of activity
posing a greater demand on the capacity of the
in this area; to a decrease in brain activity with no
auditory system. Unilateral hearing loss represents
loss of structural integrity; or to disintegration of
another example of adaptive cortical reorganization
distributed cooperative activities, that is, alteration of
in the auditory system. Sche¿er et al. (1998) report
spatiotemporal patterning of brain activity secondary
smaller response amplitudes for monaural than for
to focal damage. To assess focal brain damage, we
binaural stimulation in individuals with normal
must measure the extent of the necrotic center ver-
hearing. However, in unilaterally deaf persons, cor-
sus the surrounding area of reduced perfusion ( pen-
tical activation to monaural stimulation of the intact
umbra) and the outer edematous zones using fMRI,
ear equals the response to binaural stimulation in
Functional Reorganization of Human Cerebral Cortex 143
and we must learn what we can about the status of to improve functional capabilities of the a¤ected
function of the particular area. Using MSI and f MRI upper extremity in chronic hemiplegic patients
we can assess the functionality of areas directly (Taub et al. 1993; Taub, Crago, and Uswatte 1998;
a¤ected by a stroke and their neighboring regions Wolf et al. 1989). CIMT involves constraining the
using somatosensory magnetic fields evoked sensory una¤ected extremity and intensively shaping the
alterations (Maclin et al. 1994). movements of the a¤ected extremity, thus inducing
Alexander Luria (1963) postulated that recovery patients to greatly increase the use of the a¤ected
after brain injury depends largely on the reorganiza- extremity. Such treatment provides massively
tion of complex neuropsychological systems. Corti- enhanced and behaviorally relevant stimulus con-
cal plasticity and functional reorganization following ditions that are known to induce use-related reor-
brain lesions are thus of particular interest in neu- ganization. Kopp et al. (1999) reported substantial
rological rehabilitation. Systematic investigation of cortical reorganization of motor areas in the unaf-
functional plasticity following brain injury in animal fected hemisphere in six stroke patients after CIMT.
models strongly support this early assumption. Good EEG measurement revealed that activity (readiness
recovery is always correlated with enhanced con- potential) evoked in the ipsilateral primary motor
nectivity, whereas poor recovery is always correlated cortex by moving the a¤ected arm was greater after
with the absence of altered neuronal reorganization. than before treatment. Most likely these results re-
Furthermore, factors that stimulate functional re- flect the recruitment of intact primary motor areas to
covery, such as neurotrophins or experience, stimu- control motor function of the a¤ected hand. Using
late synaptic changes and dendritic growth, whereas TMS, Liepert and coworkers (1998) reported
factors that retard functional recovery, such as de- increased excitability of damaged motor areas after
pletion of neuromodulators, block synaptic change treatment, indicating reorganization of adjacent
(Kolb 1999). These results support the hypothesis motor areas. In both studies, therapy-induced
that neuronal plasticity underlies behavioral and changes in motor organization were accompanied by
cognitive changes and thus is the key mechanism of substantial improvements of motor function in daily
functional recovery after brain injury. It was shown life situations. It is important to note that such
in stroke patients, for example, that successive alter- improvements are rarely, if ever, achieved by other
ations in the motor cortex occurring within the first therapeutic interventions such as the Bobath
four months after the stroke were significantly re- approach or proprioceptive neuromuscular facilita-
lated to clinical improvements (Traversa et al. 1997). tion (Duncan 1997).
Although we do not know whether rehabilitation
treatments and functional reorganization interfere
Acknowledgments
with each other, and if they do, how, it appears
reasonable to expect that recovery after brain injuries
This work was supported by the German Research
can be greatly enhanced with treatment that sup-
Foundation. We are grateful to Herta Flor, Stefan
ports use-dependent adaptation of the damaged and
Knecht, Christo Pantev, and Edward Taub for many
of the related intact brain structures. Constraint-
stimulating discussions and to Lisa Green for editorial
induced movement therapy (CIMT, Taub 1994;
assistance.
Taub, Uswatte, and Elbert 2001), for example, helps
144 Annette Sterr, Thomas Elbert, and Brigitte Rockstroh
Notes
8.1 Introduction
Figure 8.2
Speech perception abilities of two groups of children for electrical stimulation alone: those born without hearing or
becoming deaf before language has developed ( prelinguistic subjects, top panel) and those becoming deaf after developing
language to sound ( postlinguistic subjects, bottom panel). Their best perception skills preoperatively and postoperatively
ranged from mere detection of sound to recognition of open sets of words and sentences that were age appropriate.
strategies F0/F2 and F0/F1/F2, and although speech the residual auditory nerve fibers (Clark 1986). With
perception was comparable between the two groups, multiple-channel stimulation, the speech frequencies
speech understanding of open sets of words in sen- are filtered into a number of frequency bands. The
tences was better for the postlinguistically deaf chil- peaks of energy or formants in the speech signal can
dren. The benefits are also greater the younger the be used to stimulate appropriate electrodes in the
child when implanted (figure 8.3). cochlea through place coding, where frequency is
recognized according to the site of stimulation in
8.1.2 Principles the brain. Sound is first filtered by the inner ear
(cochlea), with high frequencies at the basal end and
Coding of Frequency and Intensity low frequencies at the apical end (figure 8.4); the
The multiple-channel (multiple-electrode) cochlear nerve impulses (action potentials) generated from the
implant gives better speech perception than single- di¤erent frequency regions of the inner ear are then
channel (single-electrode) electrical stimulation of conveyed to the neurons in the auditory brain cen-
Learning to Understand Speech with the Cochlear Implant 149
Figure 8.4
Tuning curves ( left) show how the cochlea filters sound through the place coding of frequency, with high frequencies in the
basal turn and low frequencies in the apical turn. Diagram of auditory cortex (top right) and cochlea (bottom right) shows
how frequencies of best response of the neurons are ordered anatomically for a frequency scale to occur. SPL, sound pressure
level.
150 Graeme M. Clark
Figure 8.6
First (bottom) and second (top) formant frequencies for the syllables / ba/ (left), /da/ (middle), and /ga/ (right), highlighting the
change in the second formant frequencies.
Figure 8.8
(Left) Nucleus 24 ESPrit behind-the-ear speech processor, which activates the implanted Nucleus CI-24 Contour receiver-
stimulator; (right) child wearing the behind-the-ear speech processor.
Richards (Minimed; Merzenich and White 1980) In both cases, the speech processor extracts the ap-
systems provide respectively, analog and pulsatile propriate signals, refers the extracted message to a
stimulation to the auditory nerve fibers. More suc- map where the thresholds and dynamic ranges for
cessful than Ineraid and Minimed has been the con- stimulating each electrode for that particular patient
tinuous interleaved stimuli (CIS) strategy (Advanced are recorded. The appropriate electrical outputs for
Bionics; Medel), which not only extracts the out- each electrode are then determined and converted
puts of 6- to 8-band-pass filters, but uses, as does into a digital signal for transmission through the skin
the Speak strategy, a constant rate of stimulation by inductively coupling (radio waves) to the antenna
( Wilson et al. 1991). Results with the CIS strategy of the implanted device. Power to operate the
are more similar to those of Speak than to those of implanted section is also transmitted inductively
other multiple-channel strategies. through the skin. The implanted receiver-stimulator
then decodes the signal and transmits a pattern of
8.1.3 Engineering Implementation stimuli through the 22 electrodes to stimulate up
to 21 separate groups of auditory nerve fibers. This
The cochlear implant, as illustrated in figure 8.1, temporal and spatial pattern of stimulation conveys
consists of an external wearable component and an the speech information required for its understand-
internal implanted section. In the external com- ing in the higher centers.
ponent, a directional microphone placed in a unit The Nucleus 24 system also has telemetry (figure
above and behind the ear converts sound into elec- 8.9) that enables the voltages on the electrodes
trical energy, which is transmitted to a speech pro- in the cochlea to be transmitted externally, and the
cessor worn on a belt or in a bag hung from the pathological changes in the cochlea a¤ecting the
neck. Alternatively, Cochlear Limited has developed electrode-tissue impedance determined. In addition,
a behind-the-ear speech processor worn like a neural responses from the auditory nerve, brain stem,
behind-the-ear hearing aid (see figures 8.1 and 8.8). and auditory cortex can be recorded to help deter-
Learning to Understand Speech with the Cochlear Implant 153
Figure 8.13
Word scores for open-set Sentence Intelligibility Test (SIT) for thirteen children comparing the Multipeak strategy with the
Speak strategy at 0–3 months and 4–6 months after changeover. (Top) Sentences in quiet; (bottom) sentences in noise at a
þ15 dB signal-to-noise ratio. (From Cowan et al. 1995.) Speak scores significantly higher than Multipeak scores, according
to a one-way analysis of variance at p < 0:0:5, are asterisked.
ing speech. At least one other factor is required for significant improvement at six months after changing
speech perception, namely, language. strategies for four out of thirteen children taking the
Sentence Intelligibility Test (SIT; developed by
8.3.3 Cochlear Implants and Postdevelopmental the Clarke School for the Deaf ) sentences in quiet.
Plasticity However, the performance for M.A. deteriorated.
Note that, for one child (B.D.), there was a decrease
Postdevelopmental plasticity in older children has in scores within the initial three months after
been studied by comparing the results of changing changeover, but scores returned to similar levels for
children from the Multipeak to the Speak speech- the Multipeak strategy by six months. This trend was
processing strategies. It was not clear whether the seen for other children, but was not statistically
pattern of stimulation in the brain for one speech- significant. It suggests that a period of learning is
processing strategy could prevent the learning of required to e¤ectively use the new strategy. Note
another one that provided more information. The also that the results were similar when the children
Multipeak strategy selects up to five spectral peaks were presented speech material in quiet and in noise.
and stimulates at a rate proportional to the voicing The need for learning is illustrated in figure 8.14,
frequency. On the other hand, the Speak strategy which shows that six out of seven children tested
selects six or more spectral maxima and stimulates at were able to gain significant improvement with
a constant rate with amplitude variations conveying longer use of the Speak strategy. These findings
voicing information. Results in figure 8.13 reveal a suggest that, although children have learned to asso-
Learning to Understand Speech with the Cochlear Implant 157
Figure 8.14
Word scores for open-set Sentence Intelligibility Test (SIT) in noise at þ15 dB signal-to-noise ratio for seven children
comparing the Multipeak strategy with the Speak strategy at 6, 12, and 18 months after changeover. (From Cowan et al.
1995.) Speak scores significantly higher than Multipeak scores, according to a one-way analysis of variance at p < 0:0:5, are
asterisked.
8.3.4 Learning to Perceive Speech with Cochlear Thus the data suggest that, to learn speech, the
Implants essential sensory information must be transmitted,
and that skills at a higher processing level are
Factors A¤ecting the Perceptual Learning of required for speech perception. In studies of children
Speech using the initial speech-processing strategies, the
The factors correlated with learning to perceive open-set scores were poorer for children who were
speech in adults have been evaluated in a number of born deaf or who lost hearing before learning lan-
studies (Dowell et al. 1985; Hochmair-Desoyer and guage than for those who were postlinguistically
Burian 1985; Blamey et al. 1992, 1996), which have deaf. The final result also depended on a number of
shown that a long duration of deafness and elderly other factors: age at implantation, length of deafness,
patients are associated with poorer results. Although mode of education, and duration of implantation.
a longer duration of deafness may result in a greater
loss of neurons or their connections, loss of ganglion Rate of Perceptual Learning in Children and
cells does not appear to be the reason that duration Adults
of deafness leads to poorer results (Blamey et al. In presenting speech through electrical stimulation
1996; Nadol 1989). On the other hand, the duration of the auditory nerve in both children and adults,
of deafness is correlated with age, which is the more learning is required for maximum understanding.
probable explanation because learning is more di‰- The degree of learning over a period of twelve
cult for the elderly with a number of perceptual months for adults who had lost their hearing well
tasks. after birth and who used the inaugural F0/F2 strategy
When stimulating the auditory nerve as a whole is shown in figure 8.16. As can be seen, the mean
with an electrode placed on the inner wall of the open-set CID sentence score for electrical stimula-
middle ear, there was also a positive correlation be- tion alone increased 150% from three to twelve
tween preoperative tests of temporal processing and months postoperatively (Dowell et al. 1985).
speech scores. The ability to detect changes in rate of On the other hand, the Speak strategy reached
stimulation and gaps between stimuli appears to be higher levels and a plateau within one to two
a central function: it is important for segmenting months. By contrast, the F0/F2 strategy, which pro-
speech and processing the slow frequency changes vides less information, had not reached a plateau by
that occur in voicing. In addition, a positive rela- twelve months of implant use, which indicates that,
tionship was seen between the number of electrodes given more information with the Speak strategy, the
in use and speech perception, which highlights the adults learn more quickly than for the F0/F2 strat-
importance of multiple-electrode stimulation needed egy, and reach a plateau at an earlier stage.
for the spectral information in speech. Finally, there The rate of learning for speech perception in
was a positive correlation between the dynamic children, as evaluated by Dowell, Dettman, and
range and speech score. The greater the dynamic Barker (1998), shows some interesting changes,
range (between the threshold and maximum com- depending on the age of the child at implantation.
fortable level), the more steps in loudness that are In a study of the children at the University of
likely for presenting speech information. Melbourne’s Cochlear Implant Clinic at the Royal
Learning to Understand Speech with the Cochlear Implant 159
Figure 8.18
First and second formant frequencies for four short Aus-
Figure 8.16
tralian vowels presented to appropriate electrodes in the
Open-set Central Institute for the Deaf (CID) speech
inner ear through place coding in patient 3. Formant (F1–
scores for electrical stimulation alone over time for adults
F2) electrode separation indices are asterisked.
using the inaugural F0/F2 and the recent Speak cochlear
implant strategies.
Figure 8.19
Scores for patient 3 on the discrimination of minimal vowel
Figure 8.17 pairs of similar short durations pretraining (unshaded zones)
Postoperative changes in open-set word scores for electri- and posttraining (shaded zones). Formant electrode separa-
cal stimulation alone in children divided according to the tion indices are asterisked. (From Dawson and Clark 1997.)
age at implantation. Subjects with less than three years’
postoperative experience scored significantly higher than
those with more than five years’ experience, according to a
one-way analysis of variance at p < 0:05.
160 Graeme M. Clark
Sinusoidal gratings are one-dimensional patterns where L 0 ¼ ðL max þ L min Þ=2 is the average lumi-
where luminance varies sinusoidally along one di- nance, m is contrast, fs is spatial frequency, and f is
rection and is constant along the orthogonal direc- spatial phase.
tion (figure 9.1, panel A). A sinusoidal grating is A complex (or compound) grating would result
defined by four variables: contrast, spatial frequency, from the sum of two or more sinusoidal gratings
orientation, and spatial phase. By analogy with of di¤erent spatial frequencies. The component of
sound (where, according to the Fourier theorem, lowest spatial frequency is called ‘‘fundamental
every complex sound wave can be expressed as the component.’’ Harmonic components have spatial
162 Adriana Fiorentini and Nicoletta Berardi
Figure 9.2
(A, B, C, D) Sinusoidal luminance profiles (left) of spatial
frequency f ( panel A) and 3f with relative phase 0 ( panel
B), 180 ( panel C), and 90 ( panel D). (A þ B, A þ C,
A þ D) Luminance profiles (right) of complex gratings,
sum of f and 3f sine waves, with relative phase 0 ( panel
A þ B), 180 ( panel A þ C) and 90 ( panel A þ D).
most of these neurons undergo adaptation after- excitation and increase in inhibition have been pro-
e¤ects, resulting in a decrease of contrast sensitivity posed as possible substrates for cortical adaptation
for stimuli of the same spatial frequency and orien- (see Barlow 1990a for discussion). Carandini and
tation as those of the adapting stimulus (Ma¤ei, Ferster (1997) have suggested that the most likely
Fiorentini, and Bisti 1973; Movshon and Lennie explanation is the presence in cortical cells of a par-
1979; Albrecht, Farrar, and Hamilton 1984). The ticular tonic excitatory input (or excitatory current)
aftere¤ects of adaptation typically last several seconds that is reduced following adaptation and that slowly
after one-minute exposure to the adapting stimulus, recovers, mirroring the recovery of the response
a time course similar to that observed psychophysi- amplitude. The presence of this particular current
cally in humans. No similar decrease of contrast sen- could explain why only cortical neurons adapt to
sitivity after high-contrast stimulation is observed in contrast, whereas retinal and lateral geniculate nu-
neurons at subcortical levels. cleus (LGN) neurons do not.
The attenuation in contrast sensitivity following Thus adaptation to gratings is a striking example
adaptation and the contrast gain control of single of short-term modification in visual perception that
neurons are two aspects of plastic changes taking can be interpreted in terms of short-term plasticity in
place in the visual system. Perceptual aftere¤ects of the response of single cells or of small cell assemblies.
adaptation can be interpreted in terms of changes of The presence of plasticity at such a basic level of
response properties of a class of neurons selectively visual function (detection) responds to the necessity
tuned to the adapting stimulus. This interpretation of continuously adapting the working point of the
requires us to assume that, although a grating stimu- visual system to the most recent history of inputs to
lus may excite neurons tuned to somewhat di¤erent its neural components.
spatial frequencies, the psychophysical contrast sen-
sitivity is determined by the amplitude of the re-
9.4 Discrimination of Simple and Complex
sponse of the neurons whose sensitivity peaks at the
Gratings
spatial frequency of the stimulus. The change in ap-
parent frequency for stimuli of lower or higher spa-
Before introducing perceptual learning in the dis-
tial frequency than the adapting stimulus requires
crimination of complex gratings, let us briefly de-
that the perceived spatial frequency results from the
scribe the limits of the ability to discriminate simple
balance of response amplitude between populations
and complex gratings. Simple (sinusoidal) gratings
of neurons tuned to di¤erent spatial frequencies.
can be discriminated in terms of di¤erences in spatial
Following adaptation to a given spatial frequency,
frequency, contrast, or orientation. Under optimal
the response of the neurons tuned to that spatial
conditions, all three tasks can be performed quite
frequency is mostly depressed, and the peak of
accurately. For spatial frequencies below about 12
activity in response to a somewhat lower (or higher)
cycles/degree and for contrast three times the detec-
spatial frequency is shifted toward lower (or higher)
tion threshold, sinusoidal gratings di¤ering in spatial
frequencies (see Braddick, Campbell, and Atkinson
frequency by only 8% can be discriminated (Camp-
1978 for review). The cellular mechanisms underly-
bell, Nachmias, and Hukes 1970). On the other
ing adaptation to high-contrast patterns in cortical
hand, for pairs of gratings with spatial frequencies
neurons have long been debated; both reduction in
168 Adriana Fiorentini and Nicoletta Berardi
discrimination of these patterns has been interpreted sensitivity for these chromatic gratings is worse than
in terms of the activity of broadband channels with for luminance gratings, even if the contrast is scaled
di¤erent sensitivity profiles, either even symmetric to compensate for di¤erence in detection threshold
or odd symmetric (Field and Nachmias 1984; Ben- (Troscianko and Harris 1988; Martini et al. 1996).
nett and Banks 1987) with properties similar to those
of single neurons with sine or cosine sensitivity pro-
9.5 Perceptual Learning in the
files observed in the visual cortex of animals (figure
Discrimination of Complex Gratings Defined
9.7, panels D and E; Pollen and Ronner 1981). This
by Luminance Contrast
interpretation holds in general for other types of
complex gratings that involve the sum of two har-
The discovery of the presence of perceptual learning
monics with di¤erent phases; the findings of experi-
in the discrimination of gratings was rather seren-
ments with stimuli resulting from the sum of a large
dipitous. In the course of an experiment on pattern
number of harmonics are also consistent with the
discrimination, two observers who had done thou-
hypothesis that phase discrimination is mediated by
sands of forced-choice discriminations using vertical
two types of detectors with even and odd symmetry
complex gratings with luminance profiles such as
(Burr, Morrone, and Spinelli 1989; Martini et al.
those in figure 9.1, panels B and C, were surprised to
1996).
find that when the grating orientation was turned
For stimuli consisting of two harmonics, phase
from vertical to horizontal, they were unable not
discrimination may be worse in peripheral than in
only to discriminate the gratings, but even to see the
central vision; for eccentric patterns, it also depends
details of the patterns. Turning the grating orienta-
on the orientation of the grating: gratings oriented
tion back to vertical immediately restored the dis-
along a retinal meridian (i.e., pointing to the fovea)
crimination performance (Fiorentini and Berardi
are better discriminated than orthogonal gratings
1980). What might have happened? That grating
(Berardi and Fiorentini 1991). For gratings that
discriminability was known to be as good with ver-
consist of a large number of harmonics, however,
tical as with horizontal gratings led us to interpret
phase sensitivity is the same for central and periph-
the decline in performance as a failure to transfer the
eral vision, provided contrast is scaled according to
experience acquired in the discrimination of vertical
detection threshold (Morrone, Burr, and Spinelli
patterns to the newly presented horizontal patterns.
1989).
And, indeed, the performance of both observers
Complex chromatic gratings can be obtained by
progressively improved during further training with
superimposing in antiphase two complex gratings of
horizontal gratings until it reached a level equal to
di¤erent colors (e.g., red and green) having the same
that for vertical gratings. Because the use of a forced-
harmonic content. Sensitivity to spatial phase for
choice procedure ensured that the observers’ im-
equiluminant chromatic gratings is similar to phase
provement did not result from a change in criterion,
sensitivity for luminance gratings of the same profile,
we interpreted it as an example of perceptual learn-
provided that contrast is scaled according to detec-
ing (Gibson 1969).
tion threshold (Martini et al. 1996). Only chromatic
The discovery of perceptual learning in grating
gratings consisting of a fundamental plus third har-
waveform discrimination allowed us to infer that the
monic seem to be an exception to this rule: phase
170 Adriana Fiorentini and Nicoletta Berardi
Figure 9.12
Interhemispheric transfer of learning the discrimination of
complex gratings. Learning with gratings presented at 1
eccentricity in the left hemifield (solid circles, 240 trials
each in the first session and in the second session on the
following day) and transfer to the discrimination of stimuli
Figure 9.11 presented at the same eccentricity in the right hemifield
Selectivity of the learning process for the retinal location (open circles). The stimuli were horizontal complex grat-
of the stimuli. (A) Learning the discrimination of complex ings ( f plus 3f, as in figure 9.1, panels B and C), 5 5
horizontal gratings presented 1 above a fixation point size, with their inner edge at 1 on the right or left of
does not transfer to the discrimination of the same gratings a fixation point. (Adapted from Berardi and Fiorentini
presented 1 below the fixation point. (Adapted from 1987.)
Fiorentini and Berardi 1981.) (B) Learning the discrimi-
nation of horizontal complex gratings 4 high and 1 wide,
extending 1–2 from the vertical meridian in the left visual
the examples above, where the stimuli were sepa-
field, does not transfer to the discrimination of the same rated by at least 2 . As shown in figure 9.11, panel B,
gratings located in the adjacent region, 2–3 on the left there is no transfer of the practice e¤ects between
of the meridian. F ¼ fixation point. (From Berardi and stimuli located in the same hemifield and adjacent to
Fiorentini 1987.) each other. The specificity of perceptual learning for
the location in the visual field has subsequently been
now located in the lower hemifield. Practicing in confirmed in many other visual tasks (see chapters
this new situation brought performance back to 10–12).
nearly 100% correct in 120 trials. The same lack of There is an exception, however, as illustrated in
transfer was obtained for stimuli initially located in figure 9.12. Learning is completely transferred from
the left (right) visual hemifield (at 5 eccentricity) a location on one side of the vertical meridian to the
and then moved, upon completion of learning, in mirror-symmetric position when the distance from
the opposite visual hemifield at the same eccentricity the meridian is small (1 ). The selectivity for stimulus
(Fiorentini and Berardi 1981). The selectivity for the orientation and spatial frequency is preserved in the
stimulus location is even stricter than suggested by interhemispheric transfer of learning e¤ects (Berardi
174 Adriana Fiorentini and Nicoletta Berardi
Figure 9.15
Lack of selectivity for stimulus orientation of the e¤ects
Figure 9.14 of learning the discrimination of gratings with chromatic
(Left) Learning the discrimination of equiluminant red- contrast. (A) Learning with equiluminant red-green grat-
green ( panel A) and blue-yellow ( panel B) complex grat- ings oriented vertically (filled circles) and transfer to the
ings di¤ering for the relative spatial phase of the harmonics discrimination of the same gratings oriented horizontally
(first plus third harmonic, as for figure 9.13); (right) lack of (open circles). Stimuli as for figure 9.14. (B) Learning with
transfer to equiluminant gratings of the same waveform but blue-yellow gratings oriented at 45 (circles) and transfer
di¤erent chromaticity ( panels A and B). (From Berardi and to the discrimination of the same gratings oriented at þ45
Fiorentini in preparation.) (triangles). Stimuli: first (1 cycle/degree) plus second har-
monic, phases þ90 and 90 , presented in the lower
field, with 1 eccentricity. (From Berardi and Fiorentini
processing of luminance and chromatic contrast 1997.)
implies separate neural mechanisms that can be af-
fected independently by practice.
figure 9.15. It is clear that there is complete transfer
The next question is whether changing the chro-
of learning e¤ects when the orientation of the grat-
maticity of the isoluminant gratings from red-green
ings is rotated by 90 both for red-green and blue-
to blue-yellow (or vice versa) impairs the transfer of
yellow equiluminant gratings, contrary to what we
learning e¤ects. As shown in figure 9.14, where two
found for luminance gratings. This would imply, on
typical examples are reported, there is a substantial
the one hand, that non-orientation-selective neural
lack of transfer from red-green to blue-yellow and
mechanisms are involved in the discrimination of
vice versa. This selectivity for the chromatic chan-
chromatic gratings; yet the discrimination processing
nel is in accordance with the properties of color-
must nonetheless be located cortically because there
opponent neural channels as observed in the primate
is complete interocular transfer of learning for chro-
visual system.
matic gratings as well. On the other hand, the dis-
A rather intriguing finding in perceptual learning
crimination of symmetrical gratings di¤ering for the
of chromatic grating discrimination is illustrated in
176 Adriana Fiorentini and Nicoletta Berardi
spatial phase of their higher harmonic should require addition, the transfer of learning e¤ects between
orientation-selective neural mechanisms, which regions connected by callosal fibers, which is the
seems to rule out simply explaining the orientation only exception to the selectivity for stimulus loca-
transfer in chromatic grating discrimination in terms tion of this learning process, suggests that an essential
of color-opponent nonorientational receptive fields requirement for di¤erent groups of neurons to share
such as those found in primary visual areas of pri- the same plastic changes is the simultaneous acti-
mates. One could conceive the simultaneous use of vation of these groups of cells by the a¤erent activ-
neural mechanisms that are singularly orientation ity. This simultaneous activation does not occur
selective and the convergence of many of them for groups of neurons with di¤erent preferred ori-
on higher-order processing stages that lack orien- entations or separate receptive fields, but is present
tation selectivity. A similar hypothesis of conver- between neurons with similar preferred orientations
gence of first-order, selective units on second-order and overlapping receptive fields, as is the case for
collectors has been put forward by Morgan and callosally connected neurons.
Baldassi (1997) to explain the encoding of a defined The initial hypothesis of an exclusive involve-
texture orientation. ment of early visual areas in this perceptual learning,
put forward on the basis of the selectivity of the
learning process, is probably too naive. First of all,
9.8 Conclusions the use of a temporal forced-choice paradigm and of
a same-di¤erent discrimination task requires a short
The choice of studying a process such as perceptual time retention of the information. This may imply
learning with stimuli thitherto used to study basic some kind of working memory and therefore the
properties of the visual system proved to be particu- additional involvement of association cortical areas.
larly fruitful, enabling us, first, to demonstrate selec- Moreover, as discussed elsewhere in this book (see
tivity of the learning process for each parameter of especially Ahissar and Hochstein, chapters 14 and
a grating: orientation, spatial frequency, location, 20), locating the plastic modifications in early sen-
and luminance or chromatic contrast, and, second, to sory areas does not exclude a contribution from
compare the degree of selectivity found with the higher cortical areas. Rather, most models attempt-
selectivity for the same stimulus parameters found for ing to account for perceptual learning explicitly in-
the neural channels present in the mammalian visual clude a top-down control exerted by higher cortical
system. That the selectivity of the learning process areas. More recently, the possibility that the plastic
was very similar to the selectivity of the adaptation changes may take place exclusively at central sites has
aftere¤ects suggested initially that the plastic changes been proposed by Mollon and Danilova (1996).
underlying perceptual learning might occur at an They explain the selectivity of the learning process
early stage of visual processing, that is, in areas where by asserting that what subjects learn is to select,
neurons have small receptive fields and are selective among the many inputs to a nonselective stage, the
for the stimulus orientation and spatial frequency. incoming sets of neural signals that are most ade-
This hypothesis was reinforced by the findings on the quate to make a final decision and that are selective for
interhemispheric transfer of learning, which exhib- the stimulus parameters.
ited limits similar to those found for the callosal
transfer at the border between areas V1 and V2. In
Plasticity of Low-Level Visual Networks
Barbara Zenger and Dov Sagi
10
Abstract e¤ects, with context representing either remote im-
age parts or visual memory. Several studies (Sagi and
Although it is our everyday experience that we improve in Tanne 1994; Ahissar and Hochstein 1999; Karni and
tasks we perform repeatedly, it is in most cases unknown Bertini 1997) have shown that these (early) context
where in the brain the practice is manifested. What parts
e¤ects can be modified by practice, suggesting plas-
of the brain have been changed during or after practice?
Perhaps the most important method for answering this ticity in low-level visual networks (cf. also chapters
question is testing how specific the practice e¤ects are to 9, 11, 12).
the practiced condition and to what other tasks the prac- We introduce here a combined experimental-
tice e¤ects generalize. This chapter describes studies that theoretical approach to study context e¤ects and
document practice e¤ects in contrast-masking experiments
their plasticity. Specifically, we are trying to relate
(where observers have to detect a defined target stimulus in
the presence of defined mask stimuli) and texture seg- behavioral learning phenomena to changes in the
mentation tasks (where observers have to discern a fore- synaptic e‰cacy of simple neural networks. Such
ground region made up of one type of element from a studies are useful from two di¤erent perspectives:
background region made up of another). The observed they can reveal new information, first, about the
transfer of learning between tasks shows that practice is in type of processing that underlies the visual system,
some cases highly specific, and in others rather general. We
and second, about the rules that govern plasticity of
examine how the observed practice e¤ects and the pattern
of transfer between tasks can be explained by assuming this system.
specific synaptic changes in neural network models of the
1. Learning data constrain models of neuronal processing.
primary visual cortex.
Recording data from observers at di¤erent stages of
practice is like obtaining information about di¤erent
states of the same perceptual system. A neuronal
10.1 Introduction model with a certain architecture that accounts only
for one state of practice but not for another can be
During the first stages of visual processing, local im- rejected. Thus, even for the researcher whose pri-
age properties, such as intensity, orientation, color, mary interest is not to understand how practicing a
and spatial frequency, are extracted in processes that certain task can lead to a performance improvement,
operate in parallel over the whole visual field. The it may be worthwhile to study learning, as a way to
di¤erent feature detectors can be viewed as operating create ‘‘behavioral variants.’’ The goal is to develop a
independently only to a first approximation; as a model that naturally accounts (by simple parameter
result of the significant interactions observed, some change) for all stages of practice while being at the
features are enhanced, whereas others are suppressed. same time biologically plausible and as simple as
These mechanisms can be interpreted as a means possible.
by which the visual system mediates simple context
178 Barbara Zenger and Dov Sagi
Figure 10.1
Stimuli patterns used in the experiments reviewed here were presented for 90 msec at the center of a computer screen,
which was otherwise gray. These contrast modulation patterns were generated by modulating the amplitude of a sine wave
grating with a two-dimensional Gaussian window (Gabor 1946). Each mask configuration was shown twice: once without
Plasticity of Low-Level Visual Networks 179
2. Learning data can suggest mechanisms that govern specific unit in the primary visual network. (Such a
synaptic plasticity. Why have the changes in network unit is thought to correspond to several neurons in
connectivity occurred in precisely the way they the cortex.) By measuring the contrast of the signal
have? Why have the e‰cacies of some connections that is required for the observer to detect the stimu-
increased, whereas others remained constant or lus, one obtains an indirect measure of this unit’s
decreased? The goal is to discover a set of learning sensitivity. In subsequent experiments, the detection
rules that can successfully predict how performance threshold of the target is measured in the presence of
changes after practice of a specific task and how other stimuli (composed of Gabor patches) called
other tasks will be a¤ected by these synaptic mod- ‘‘masks,’’ which will usually either enhance or sup-
ifications. For our approach to be successful, our press detection of the target (depending on mask
behavioral paradigm should be simple and well configuration and mask contrast). These sensitivity
understood in terms of cortical processing. A good changes in the target unit can then be explained by
candidate for a simple model system of neuronal assuming specific types of interactions between mask
processing is the primary visual cortex (area V1), and target units. Facilitation points to excitatory
where cells are rather stimulus specific (e.g., location, interactions from the mask unit to the target unit,
orientation, scale), interconnected by a rich local whereas suppression points to inhibitory interactions.
network (Gilbert 1993). Both suppression and facilitation are a¤ected by
perceptual learning.
10.1.1 Psychophysics
10.1.2 Modeling
Experimentally, these local circuits can be studied
with a variety of psychophysical paradigms, such as The model emerging from these studies is one that
selective adaptation (Blakemore and Campbell 1969), consists of perceptual filters, which interact with each
subthreshold summation (Kulikowski, Abadi, and other via plastic excitatory and inhibitory inter-
King-Smith 1973), and contrast masking (Campbell actions. Each of the perceptual filters assumes a re-
and Kulikowski 1966, Legge and Foley 1980; ceptive field covering a small portion of the visual
Wilson and Humanski 1993). Here we use results field, selective for orientation and size (spatial fre-
from contrast-masking experiments (Zenger and quency; Wilson and Wilkinson 1997). Excitatory or
Sagi 1996; Dorais and Sagi 1997) to study proper- inhibitory interactions, or both, take place between
ties of the primary network units, their interactions filters with spatially overlapping receptive fields that
and plasticity. In these experiments, observers are are tuned to di¤erent spatial frequencies (Tolhurst
required to detect a briefly presented target stimulus. and Barfield 1978) and di¤erent orientations (Foley
The target is a Gabor patch (see figure 10.1, panel 1994; Zenger and Sagi 1996). In addition, there are
A), a stimulus that is designed to selectively excite a interactions between filters with similar orientation
target (left ), once with a vertical target added (right). (A) Just-noticeable di¤erence ( JND): vertical mask. (B) 45ðþþÞ: two
mask components tilted by 45 clockwise and counterclockwise, respectively; both components had ‘‘positive phase’’ (they
were white in the center). (C) 45ðþÞ: mask components were as in panel B, only one component had ‘‘negative phase’’ (it
was black in the center). (D) 30ðþÞ: a single mask component tilted clockwise by 30 .
180 Barbara Zenger and Dov Sagi
during practice, but also on the task that was 1995; Tanaka and Sagi 1998; Lee, Koch, and Braun
attended by the observer (Ahissar and Hochstein 1997) and its ‘‘readout,’’ these operations are ex-
1999; cf. chapter 14). That learning (with the same pected to be acquired on a short time scale of a
stimulus) can go in di¤erent directions depending on few trials (Karni and Sagi 1993; Karni and Bertini
the task employed means that the task e¤ect on 1997). In other words, we assume that performance
learning cannot simply be seen as a switch that is set in the tasks considered here is limited exclusively by
into ‘‘on’’ position when the task is attended and into the low-level visual network, irrespective of whether
‘‘o¤ ’’ position when the task is not attended. In observers are naive or practiced. (Of course, whether
texture learning (Karni and Bertini 1997), where a observers reach these limits or not may depend on
target consisting of several oriented lines is to be whether they fully attend the task or not.)
detected in an array of lines having a di¤erent ori-
entation, performance improvement for a specific
10.2 Short-Range Interactions
target was found to depend on the relevance that this
specific target had in the practice sessions. Post-
We first describe experiments where the target was
learning performance may show (1) improvement
masked with Gabor patches presented at the target
when the tested target is the one practiced (Karni
location, but with varying orientations. These masks
and Sagi 1991), (2) no change in performance rela-
presumably excite units which are located within the
tive to the initial value when the tested target pre-
same hypercolumn as the target unit, and they thus
sented during practice has no relevance to the task
probe cortical short-range connectivity.
(Karni and Sagi 1995), and (3) reduced performance
Our masks typically consist of two components:
when the tested target served as a distractor during
one tilted counterclockwise and the other clockwise
practice (Tanne and Sagi 1995). Such results imply
with respect to the target. Each component can
that simply looking at the stimulus does not lead to
either have positive polarity (i.e., the center of the
learning: for learning to take e¤ect, we need to act
even-symmetric Gabor patch is white) or negative
on the stimulus, and the way we act determines the
polarity (where the center of the even-symmetric
learning path.
Gabor patch is black). In the experiments described
here, target polarity was always positive, and target
10.1.4 Working Hypothesis
orientation was either horizontal or vertical.
To describe a specific mask configuration, we use
Changes made in the low-level visual network to
the following terminology: the tilt angle (with re-
account for stimulus-specific learning e¤ects are pre-
spect to the target) is followed by the polarities of
sumably task specific, and would not occur when
the two components. The notation ‘‘30(þþ),’’ for
stimuli were viewed only passively, indicating in-
instance, refers to a mask configuration where both
volvement of high-level processes. We assume here
mask components are of positive polarity, and they
that these high-level processes a¤ect only learning
are tilted with respect to the target by þ30 and
(i.e., how the system moves from one practice state
to another), but not the performance level at any 30 . The notation ‘‘45(þ)’’ describes a mask con-
given state of practice. Although attentional pro- figuration where the component that is tilted coun-
cesses may modulate network activity (Ishai and Sagi terclockwise (by 45 ) has positive polarity, whereas
the component tilted clockwise (also by 45 ) has
182 Barbara Zenger and Dov Sagi
negative polarity. In some conditions, only one mask that the target unit does not respond to low-contrast
component was used, tilted either counterclockwise, stimuli, but responds only if the stimulation exceeds
referred to as ‘‘ðþkÞ’’, or clockwise, referred to as a defined threshold (but see Pelli 1985 for an alter-
‘‘ðkþÞ’’. When the mask orientation corresponds to native explanation using uncertainty concepts). If the
the target orientation, the task becomes a simple con- mask provides excitatory input to the target unit it
trast discrimination task. The corresponding masking brings this unit closer to threshold, and thus increases
curve is called the ‘‘just-noticeable di¤erence’’ ( JND) its sensitivity. The magnitude of excitatory input
curve. (For examples of di¤erent mask configura- from a specific mask can be estimated from the lo-
tions with and without target see figure 10.1.) cation (on the log mask contrast axis) of the dipper
section with negative slope.
10.2.1 The Dipper Function Note that one can distinguish two kinds of exci-
tatory input from the mask to the target unit: when
In typical masking experiments, the target’s detec- the mask is presented within the receptive field of
tion threshold is plotted as a function of mask con- the target it provides direct input, whereas, when the
trast, with mask contrast referring to the sum of the mask stimulates a unit that in turn stimulates the
contrasts of the di¤erent mask components. (All target unit it provides indirect input.
other mask parameters such as orientation, size, and
the like are kept constant.) The classical finding is Performance Decay with Increasing Mask
that, with increasing mask contrast, thresholds first Contrast
decrease, reach a minimum, and then increase, re- This regime corresponds to the Weber-Fechner
sulting in a dipper-shaped function (Legge and Foley law, which describes human sensitivity to changes of
1980). The two parts of the dipper function seem to sensory input (DI=I ¼ constant, where I is the sen-
reflect two di¤erent types of interactions. The initial sory input magnitude and DI is the JND), predict-
performance improvement is attributed to excitatory ing a slope of one on log-log coordinates. Contrast
input from the mask to the target unit, whereas the masking produces slopes smaller than 1 (see figures
performance decay for further increasing mask con- 10.3 and 10.4), thus presenting a deviation from the
trast is attributed to inhibitory interactions. Figures classical sensory law. Performance decay may be the
10.3 and 10.4 depict several examples of masking result of response saturation (Wilson 1980), inhib-
curves. Dipper functions for di¤erent mask config- itory network interactions (Heeger 1992; Foley
urations (i.e., masks with di¤erent orientations and 1994), or both (Wilson and Humanski 1993). Net-
polarities) di¤er in location and magnitude of the work interactions can be implemented as divisive
dipper, but the slopes of the two sections are usually inhibition from units that surround the target unit.
stable (when plotted on a log-log scale; see also Such a mechanism normalizes a unit’s response with
Zenger and Sagi 1996). respect to the average activity in this region, and may
thus be an important means to avoid response satu-
Performance Improvement with Increasing ration (Heeger 1992). The divisive input reduces the
Mask Contrast gain of the target unit and thus decreases its sensi-
The explanation of performance improvement with tivity. With increasing mask contrast, the inhibitory in-
increasing mask contrast is strongly related to the put becomes stronger and stronger, and performance
concept of a response threshold. The basic idea is deteriorates more and more. The magnitude of in-
Plasticity of Low-Level Visual Networks 183
Figure 10.3
Masking curves for observer I. E. in conditions with di¤erent mask configurations, before and after practice. The mask
contrast as well as the detection thresholds are presented on a normalized log scale, that is, the values shown represent the
logarithm of the ratio of specific contrast values and the detection threshold; hence zero denotes identical contrast values of
both stimuli—a ratio of 1. Masking curves di¤er in the magnitude and mask contrast value of the maximal dip (sometimes
there may be no dip at all). After practice, there was a notable decrease in suppression in the 45ðþþÞ condition (right); the
practice e¤ect was smaller in the 15ðþþÞ condition (middle); and there was no e¤ect on the just-noticeable di¤erence ( JND)
curve (left), even though long-term practice lasted for almost forty sessions in each of the conditions.
hibitory input from a specific mask configuration can detection threshold in the absence of masks) and the
be estimated from the location (on the mask contrast initial section of the dipper function are only slightly
axis) of the dipper section with positive slope. a¤ected by practice, if at all (Zenger and Sagi 1996).
Examples are shown in figures 10.3 and 10.4. To
10.2.2 Practice E¤ects characterize the e¤ect quantitatively, we estimate the
suppression threshold, which is the mask contrast at
When an observer repeats this type of masking ex- which masks start to have a negative e¤ect on target
periments for several sessions, the dipper function detection (i.e., the point where the dipper crosses
changes in a characteristic way. The most obvious the baseline) normalized by the target’s detection
change is a decrease in suppression, that is, the threshold. An interesting observation is that even
masking curve section with positive slope is shifted though for almost all masking conditions suppression
to the right (along the mask contrast axis), whereas thresholds increased with practice, such a practice
the slope is unchanged. The absolute threshold (i.e., e¤ect was never observed when the mask and target
184 Barbara Zenger and Dov Sagi
Figure 10.4
Masking curves for observer H. B. in conditions with di¤erent mask configurations, before (solid curves) and after (broken
curves) practice, actual (left) and modeled (right). Mask contrasts as well as detection thresholds are presented on a normalized
log scale. A strong decrease in suppression was observed in both conditions. In some cases, masks that led to a suppression in
the naive state produced facilitation after practice. (See section 10.2 for a description of the model.)
parameters were the same, that is, when the pattern synapse, but there are several synaptic modifications
discrimination task reduced to a contrast discrimina- which all lead to in a decrease in inhibition. In the
tion task (see figure 10.3; Dorais and Sagi 1997). following, three simple models are briefly described
How can we explain the decrease in suppression where learning is attributed to change in the e‰cacy
with the proposed neural network model? The re- of a single synapse (see figure 10.5). Each of these
gion with positive slope is a¤ected primarily by the models makes distinctive predictions on how prac-
inhibitory interactions, suggesting that the inhibitory tice would transfer between di¤erent conditions.
input from the mask on the target unit decreases Clearly, learning might be more complex, and sev-
during practice. Within the framework of the neural eral synapses might change their e‰cacy simulta-
network model considered here, the inhibition from neously, but it seems worthwhile to first explore the
the mask on the target is not determined by a single simple cases. In the following, units tuned to the
Plasticity of Low-Level Visual Networks 185
Figure 10.6
Psychophysical suppression thresholds ( left) estimated from the data of Dorais and Sagi (1997), together with model simu-
lations (right). (Note that the higher bars correspond to the weaker masking e¤ects.) The labels on the x-axis denote which
task was practiced before a specific data set was obtained. (A) Practice of the 30ðþþÞ condition followed by practice of the
30(þ) condition. Practice with the compound mask did not transfer to the individual components, but consequent practice
with one component transferred to the other component. (B) Practice of the 45ðþþÞ condition with a horizontal target,
followed by practice with a vertical target. When switching target orientation, the suppression threshold returns to the naive
level, indicating no transfer. (C) Practice of the 45ðþþÞ did not lead to a good performance in the 45(þ) condition, being
suggestive of a no-transfer situation. (Unfortunately prepractice data are missing for the 45(þ) condition.) (D) Practice in
the 45(þ) condition led to an improvement in both the 45(þ) and 45(þ) conditions.
Plasticity of Low-Level Visual Networks 189
relevant and that these interactions change during overlapping oriented filters with excitatory and in-
practice. Such e¤ects of mask interactions on target hibitory interactions. Similar interactions were found
detection cannot be incorporated into simple feed- between spatially nonoverlapping filters (Polat and
forward models. The lack of transfer from practice of Sagi 1993; Polat and Sagi 1994a; Wilson and Wil-
the compound to its components was one of the kinson 1997). These experiments involve detection
major motivations to adopt a feedback architecture. of a Gabor signal in the presence of flanking high
There is a second transfer result that has direct contrast masks (figure 10.7, panel A).
consequences for possible architectures. Dorais and
Sagi (1997) describe one observer who practiced 10.3.1 Spatial Interactions
detection of a horizontal target in the presence of a
mask consisting of two components. The polarities When the target-to-mask distance is varied, two
of all three Gabor stimuli (target component and e¤ects can be observed: increased threshold (sup-
two mask components) were positive (white in the pression) with small separations, and decreased
center). After practice, transfer was tested in a task thresholds (facilitation) with larger separation (figure
where the polarities of all three stimuli were re- 10.7, panel B). When using high-contrast masks, the
versed. The absence of transfer implies that the location of the transition between suppression and
model cannot contain only phase-insensitive units facilitation changes with the Gabor scale used
(mimicking typical complex cells) but that there (wavelength, l) and is at about 2l (Polat and Sagi
have to be phase-sensitive units (mimicking typical 1993). Interactions within the range of 2l can be
simple cells). Otherwise, performance in the two viewed as short-range interactions, mainly within
tasks (detection of negative target in the presence of the size of a receptive field (see section 10.2), with
negative mask versus detection of positive target in e¤ects being contrast and phase dependent (Zenger
the presence of positive mask) would be mediated by and Sagi 1996). The short-range lateral e¤ects reflect
the very same units, which are blind to the di¤erence a balance between lateral inhibitory interactions and
between the two tasks, and the observed ‘‘indepen- excitatory interactions, with inhibition taking over
dent variation’’ of the two performance levels would at high contrast levels (Sagi and Hochstein 1985;
not be understandable. Furthermore, it is implied Stemmler, Usher, and Niebur 1995; Polat et al.
that the modifications that occur during practice af- 1998; Somers et al. 1998; Bonneh and Sagi 1999).
fect these phase-sensitive cells; otherwise, an inde- Because we assume that the external input to a net-
pendent variation of the performance levels in the work subunit can be represented as linear integration
two tasks would not occur. Note that it would be over some retinal receptive field (linear filters), and
comparably di‰cult to obtain such strong architec- connections between subunits as nonlinear inter-
tural constraints without the analysis of learning data. actions, we expect processes within the subunit in-
tegration area (receptive field) to have a spatial
phase–dependent component, whereas long-range
10.3 Long-Range Spatial Interactions
interactions may show phase independence (Zenger
and Sagi 1996; Bonneh and Sagi 1998).
In section 10.2, we presented a framework for
analyzing orientation masking data, using spatially
190 Barbara Zenger and Dov Sagi
Figure 10.7
(A) Stimulus used for exploring lateral interactions, with two high-contrast masks (Gabor signals) flanking a low-contrast
target. (B) Data from the lateral masking experiments, before and after practice. (Redrawn from Polat and Sagi 1994b.)
Observers detected a Gabor target flanked by two high-contrast colinear Gabor signals at di¤erent distances. Contrast de-
tection thresholds relative to absolute threshold (no mask) are plotted as a function of target to mask distance, using Gabor
wavelength units. Data are shown for two observers (R. M. and G. H.) and for a model network assuming activity propa-
gation through lateral excitatory connections with local normalization. (From Sagi 1996.) Observers show an increased range
of enhancement after extensive practice. The model emulates practice by increasing excitatory synaptic e‰cacy and thus
increasing the range of signal propagation through the excitatory lateral connections. Note that suppression is not a¤ected by
practice in these experiments and simulations.
10.3.2 Network Architecture in these low-level networks, and limits the ‘‘learn-
able.’’
Long-range spatial interactions are configuration de-
pendent (figure 10.8; largest facilitation is obtained 10.3.3 Learning
with cooriented and coaxial target and masks or with
cooriented but parallel target and masks; Polat and Long-range spatial interactions are experience de-
Sagi 1994a). In terms of the model network pre- pendent. Both excitatory and inhibitory e¤ects may
sented here (figure 10.2), these results indicate that change with practice, depending on stimuli used.
excitatory connections (Ei ! Ej ) take place between As a general rule, it seems that frequently activated
excitatory subunits corresponding to neighboring connections increase their e‰cacy with practice
spatial locations, only in directions along, and or- (though depending on some task-dependent gating;
thogonal, to the subunit’s preferred orientation. In- see section 10.1).
hibitory connections (Ei ! Ij ) are isotropic (but see
section: 10.3.3 on nonisotropic modulation of con- Learning Excitatory E¤ects
nections’ e‰cacies). This result imposes a strong Observers practicing the lateral masking task, with
constraint on possible interactions that can take place target flanked by two high-contrast masks at varied
Plasticity of Low-Level Visual Networks 191
Figure 10.8
Stimuli used for exploring performance limits imposed by network connectivity. (From Polat and Sagi 1994a.) Lateral
facilitation is greatest with colinear configurations, independent of local or global orientation. Parallel configurations (i.e.,
local orientation orthogonal to global orientation) produced lateral facilitation, whereas locally tilted configurations (i.e.,
local orientation diagonal to global orientation) failed to produce facilitation even after extensive practice (Polat and Sagi
1994b); lateral inhibition a¤ected all configurations (Polat and Sagi 1994a).
distances, show an increased range of facilitatory 10.8), but not along other directions (Polat and Sagi
e¤ects. Depending on the range that is practiced the 1994b).
interaction range can reach up to seven times the
receptive field size (Polat and Sagi 1994b). This Learning Inhibitory E¤ects
learning is highly specific for basic visual features Whereas masks at distances of 3l and more seem
such as orientation, location, spatial frequency, and eye to probe mainly excitatory interactions, masks pre-
(Polat and Sagi 1995). To account for the increased sented at distances of 2l and less appear to involve
facilitatory range, there seem to be two plausible also inhibitory interactions (see contrast and range
alternatives: practice causes a selective increase either e¤ects in section 10.3.1). Practicing a lateral masking
in the e‰cacy of direct long-range interactions be- task with multiple high-contrast flankers placed at 2l
tween mask unit and target unit (Em ! Et ) or in the intervals (see figure 10.9) was found to increase the
e‰cacy of synaptic transmission between neighbor- range and e‰cacy of these inhibitory interactions
ing subunits (Ei ! Eiþ1 ), positioned between mask (Adini, Sagi, and Tsodyks 1997). Results from these
and target units. Experimental evidence supports the experiments are presented in figure 10.10 for chains
latter, pointing to the existence of a feedback net- of colinear Gabor signals (figure 10.9, panel B). For
work in which local interactions enable an e‰cient unpracticed observers, increasing the number of
propagation of neuronal activity. flankers beyond two or four (one or two on each
Propagation is not isotropic (see section 10.3.2), side) has no e¤ect on contrast threshold. For prac-
but is constrained by existing connections. Learning ticed observers, increasing the number of flankers has
studies have shown learning e¤ects of facilitatory a nonmonotonic e¤ect on threshold: adding flankers
interactions along two cardinal directions (coaxial cancels the facilitation, and adding still more flankers
and parallel relative to local orientation; see figure restores some of the facilitation.
192 Barbara Zenger and Dov Sagi
Figure 10.9
(A, B) Stimuli used in chain lateral masking experiments: parallel configuration ( panel A); colinear configuration ( panel B).
(From Adini, Sagi, and Tsodyks 1997.) Target is flanked by a total of ten Gabor signals: five on each side, spaced 2l apart.
These results indicate the development of lateral tion tasks improves with increasing texture density
inhibition between neighboring units that respond (Nothdurft 1985; Sagi and Julesz 1987; Sagi 1990),
to the high-contrast flankers, possibly involving pointing to a performance limit set by network
redistribution of mask gain control (section 10.2). connectivity range. Karni and Sagi (1991), using
These e¤ects were found to be configuration spe- dense textures composed of line elements with three
cific: colinear stimuli (figure 10.9, panel B) showed tilted lines serving as a target and 357 horizontal lines
little indication for inhibition at the initial phase of as the background (figure 10.11), found large per-
practice (figure 10.10), whereas parallel stimuli (fig- formance improvement with practice. In a later
ure 10.9, panel A) seemed to successfully utilize study (Karni and Sagi 1993), they found that some
existing strong inhibitory interactions. Practice had fast performance improvement occurs within the
the e¤ect of equalizing inhibitory e¤ects along these first practice session, followed by a slower learning
two cardinal directions. that takes a few sessions to converge; that, once
In addition to the increase in lateral inhibition, a learning reaches an asymptotic level, it can last for a
second practice e¤ect was observed in these experi- long time without further practice, at least for two
ments: practice increased facilitation of target detec- years; and that these di¤erent time scales do not seem
tion by neighboring flankers (figure 10.10), probably to reflect a time-dependent decrease in e¤ectiveness
the result of decreased inhibitory e¤ects from masks of a unique mechanism, but rather correspond to
to target (see section 10.2 on redistribution of target distinct mechanisms, operating at di¤erent levels of
gain control). brain processing, as implied by the di¤erent selec-
tivity properties of these two temporal regimes.
10.3.4 Applications to Texture Learning During slow learning, memory traces may take a
few hours to consolidate, with consolidation being
In texture segmentation tasks, observers detect an enhanced during night sleep, with the di¤erent sleep
image region (target, foreground) that is di¤erent stages having selective e¤ects (Karni et al. 1994;
from the background. When target and background Stickgold et al. 2000) on the memory traces.
di¤er in some distinctive basic feature (e.g., orien- The main finding of Karni and Sagi (1991) is that
tation), segmentation is said to be ‘‘e¤ortless’’ and learning is specific for some basic visual features,
preattentive ( Julesz 1981), and performance is lim- such as orientation, location, and eye. Monocularity
ited by the low-level visual network (Rubenstein of learning implies an anatomical learning site below
and Sagi 1990). Performance on texture discrimina- or at the site where the images from the two eyes
Plasticity of Low-Level Visual Networks 193
remains constant. Only those background units that crease within groups and decrease between groups.
surround the target become more active. This pre- Although network states at di¤erent stages of prac-
diction is supported by recent recordings from tice can be characterized by the suggested segmen-
monkey visual cortex (Bertini et al. 1995, 1996) tation process, the rules governing the transitions
showing constant response through practice of single between the di¤erent network states are not obvi-
units that respond to texture background elements, ous. One question left open here is how the system
but increased response in units that respond to knows which units correspond to the target and
background elements in the neighborhood of a tar- which units correspond to the mask (or, in general,
get. Second, the mechanism seems consistent with which unit corresponds to which functional group).
the finding that (slow) texture learning is specific for There may be several plausible answers to this ques-
background orientation but not for target orienta- tion; two possibilities are briefly described.
tion (Karni and Sagi 1991). It would further be
1. The system may use high-level information. In the
consistent with the finding that conditions involving
experiments described here, observers are well in-
di¤erent background orientations can be learned
formed about what is the target they are supposed to
independently of each other (Ahissar et al. 1998;
detect, and what is the mask (and even if they were
Karni and Sagi 1991), that is, when observers prac-
not told, they might still be able to figure out on a
tice several tasks, these practice e¤ects do not need
cognitive level what is the task). This type of high-
to interfere with each other. Of particular interest
level information might be used to direct synaptic
here are conditions where target and background
modifications that result in an e¤ective segregation
orientations are swapped, thus creating an apparent
of the two patterns.
conflict between two learned conditions. Our net-
work, after practicing these two conditions, produces 2. The system may use low-level information. The
equally good performance in these two conditions. activity of units belonging to di¤erent functional
groups have (at least to some degree) uncorrelated
activities, whereas the activities in di¤erent units
10.4 Conclusions belonging to the same functional group are highly
correlated. Knowing, for instance, the activity in one
Perceptual learning appears to produce significant of the mask units completely determines the activity
changes in networks of early vision, changes that in other mask units, but does not say much about the
modify the e¤ective connectivity of the network. activity in the target unit (i.e., whether the target is
The type of modification taking place presumably present or absent). Note that these correlations can-
depends on stimulus and task. We suggest some not be estimated within a single trial, but require
simple, certainly simplified, rules that were devel- that information be integrated across many trials or
oped to account for contrast detection data, but several blocks. Presence or absence of correlations
have, in addition, some more general flavor. may allow the system to determine the units that
When practicing visual tasks that include several belong to di¤erent functional groups.
objects—some of them targets—inhibitory e¤ects
are redistributed to achieve independent gain control These learning rules are not necessarily specific
within functional groups: inhibitory interactions in- for low-level vision, but may apply also to other
domains where objects need to be segmented. The
196 Barbara Zenger and Dov Sagi
Acknowledgments
Figure 11.3
Figure 11.2 Stimulus (top) and photoreceptors (bottom) for projection of
Stimulus (top) and schematic photoreceptors (bottom) for a point or line situated at marginally di¤ering positions in
projection of a point light source in the outer world onto the outer world onto the foveal receptor mosaic. (a) Posi-
the retina. (a) Optics of unlimited resolution stimulates only tion exactly above the middle of the photoreceptor, where
one photoreceptor. (b) Optics with a point spread func- the middle photoreceptor is stimulated most strongly
tion similar to the human eye’s optics, stimulates several (darkest shading) and both its neighbors equally less stim-
photoreceptors. ulated (medium gray). (b) Position shifted slightly to the
left, where the neighbors are stimulated asymmetrically.
Figure 11.4
Schematic reconstruction of an arbitrary function accord-
ing to the sampling theorem. The function (here a sine- Figure 11.5
wave) is sampled either more than twice per period (arrows Parallel versus serial detection of vernier o¤sets. (a) Vernier
from below), allowing perfect reconstruction, or about o¤set to the left is immediately perceived among straight
once per period (dotted arrows from above), not allowing distractors: it ‘‘pops out.’’ (b) Vernier o¤set to the left is not
correct reconstruction. easily found among verniers o¤set to the right but requires
time-consuming, serial search.
quency contained in the signal (figure 11.4). The
requirements of the sampling theorem are fulfilled in crimination thresholds that are below a photore-
the human visual system. The resolution of the eye, ceptor diameter: hyperacuity is thus not just an
as mentioned above, is limited: the highest spatial achievement of highly trained observers. A second
frequency reaching the retina contains around 60 argument is that no scrutiny is required for vernier
cycles per degree. Because, however, the spacing of judgments (i.e., discriminations between aligned
foveal photoreceptors is around 35 arc sec, all but versus misaligned vernier elements): they can be
the highest spatial frequencies present in the retinal made in parallel over the visual field. If sixteen ver-
image are sampled twice per period. Thus the lim- nier stimuli are presented simultaneously, as in figure
ited resolution of the eye’s optics ensures that the 11.5, panel a, we immediately spot the one that is
neuronal part of the visual system receives the infor- o¤set. This feat is by no means trivial; it requires
mation necessary to determine the exact localiza- detectors in the hyperacuity range to be present for
tion of visual features beyond the grain of foveal a great number of visual field positions, so that the
photoreceptors. visual field can be processed in parallel regarding the
feature ‘‘o¤set’’ versus ‘‘straight.’’
Consider, in this regard, figure 11.5, panel b.
11.3 What Is Hyperacuity Good For? Again, we have to find the vernier whose lower
element is o¤set to the left. However, opposite to
On the other hand, perhaps vernier acuity is an panel a, all the distractors are not straight, but o¤set
unnatural ability of human observers, present only to the right. Obviously, our visual system does not
under laboratory conditions for highly trained ob- have parallel detectors in the entire visual field that
servers, and having no bearing on normal life. A first allow us to discriminate between o¤sets to the left
argument against this view is the observation that versus o¤sets to the right, so we have to perform a
even untrained observers will usually achieve dis- serial search and look through all the stimuli se-
Learning to Perceive below Foveal Photoreceptor Spacing 201
quentially until we find the target—at least if ori- Hyperacuity is an important ability of the brain
entation cues are masked (Fahle 1991). And, indeed, used for analyzing the visual world, for example, in
search times for detecting an o¤set target among making stereoscopic depth estimates from the min-
straight distractors do not strongly increase with the ute di¤erences between the images of the same
number of elements displayed simultaneously (figure object cast into the two eyes. But how can the brain
11.6, panel a), even if the vertical gaps between the improve its ability to discriminate between vernier
upper and lower vernier elements are randomly o¤sets? One possible way that springs to mind is by
varied in size. But finding a target o¤set to the left adjusting receptive field size, as shown in figure 11.7.
among distractors o¤set to the right and even to find Many receptive fields in the visual cortex consist of
a straight target among o¤set distractors requires elongated regions, stimulation of which leads to an
search times that increase almost linearly with the activation of the cell, with side bands whose activa-
number of elements displayed simultaneously ( panel tion leads to inhibition of the same neuron (see fig-
b). Similar results were obtained for the search for a ure 11.7). These receptive fields have all kinds of
target curved to the left among distractors that were orientations and cover the visual fields with a fine-
straight versus among distractors curved to the right. grained layer. As can be seen in figure 11.7, a cell
The reaction times in the first case were virtually with a receptive field oriented vertically will not be
independent of set size, whereas they increased able to discriminate between a straight and an o¤set
steadily with the number of elements in the second. vernier stimulus, as long as the vernier is centered on
The human visual system thus obviously devotes a the cell’s receptive field. But nearby cells, whose
large number of detector neurons to the task of dis- receptive field centers are shifted somewhat to the
criminating between straight and nonstraight line side compared to the first neuron, will be able to
elements. This discrimination must be an important discriminate between these two stimuli: the o¤set
one, while the discrimination between the directions stimulus will activate part of the inhibitory surround.
202 Manfred Fahle
Figure 11.8
Improvement through training in a vernier discrimination task. (Left) Mean number of correct responses by twelve observers
increased over the training period of twenty blocks with constant stimulus orientation, but improvement did not transfer to
the new orientation when the stimulus was rotated by 90 . (Right) For six control observers, stimulus orientation was con-
stant, and no drop in their performance occurred. (After Poggio, Fahle, and Edelman 1992.) Vertical bars indicate standard
errors of means.
who had trained with horizontal stimulus orientation This result clearly indicates that the improvement
now saw vertical stimuli, and vice versa. in vernier discrimination obtained through training
Although there were large interindividual di¤er- is specific for stimulus orientation, at least for the
ences, mean performance of all observers fell o¤ relatively fast component of perceptual learning that
sharply at the time of stimulus rotation. Initial per- takes place within the first half hour or so of training.
formance for the new orientation was even below To assess whether this orientation specificity is due
the baseline level of the same observers at the start to some form of fast switching processes or whether
of the experiment, that is, for the first orientation. it also occurs for long-term learning, we performed a
During the following hour, average performance similar experiment with another group of twelve
of all observers improved steadily, especially during observers.
the first half hour, up to the level obtained for the In this control experiment, discrimination thresh-
first orientation. For a control group of another six olds, defined as the o¤set size yielding 75% correct
observers, stimulus orientation remained constant responses, were measured rather than percentages of
throughout the experiment (figure 11.8, panel b). correct responses as in the previous experiment. The
There was no dip in performance for this group of results are plotted in figure 11.9 and show that this
observers. group of twelve unexperienced observers yielded
204 Manfred Fahle
Figure 11.9
Orientation specificity of perceptual learning with vernier stimuli. Average thresholds of twelve observers improved over the
five-hour period of learning, but increased sharply when the stimulus was rotated by 90 . (After Fahle and Edelman 1993.)
Vertical bars indicate standard errors.
thresholds around 13 arc sec initially, roughly corre- reached with the first orientation. Thus we can es-
sponding to one-third of a foveal photoreceptor tablish that improvement of performance in a vernier
diameter. Again, six observers initially trained with discrimination task is highly specific for stimulus
horizontal stimulus orientation, whereas the re- orientation and that both the fast and the slow phase
maining six started with vertical orientation of the of learning are specific for the orientation of the
stimulus. Training continued in one-hour sessions vernier stimulus, as is the case for grating stimuli (cf.
usually on subsequent days, for five hours altogether. chapters 9, 10).
Thereafter, stimuli were rotated by 90 , as in the first
experiment. As a result, thresholds rose dramatically,
11.6 Position and Eye Specificity of
to more than 15 arc sec. Thus the rebound over-
Perceptual Learning
shoots: initial performance for the second stimulus
orientation is clearly inferior to that for naive
In the next experiment, we tested whether im-
observers as tested with the first orientation. In other
provement in the vernier discrimination task
words, training a vernier discrimination with a stim-
achieved at one visual field position would transfer
ulus in one orientation improves performance for
that orientation, while clearly impairing perfor- to another visual field position. Another group of
eight naive observers performed a vernier discrimi-
mance for stimuli at right angles to that orientation.
nation task at 10 eccentricity (figure 11.10). The
Over the next five hours of training, observers
gradually improved performance and finally obtained eight positions to be tested were arranged, at con-
levels of performance comparable to the final levels stant distances from one another, on an imaginary
Learning to Perceive below Foveal Photoreceptor Spacing 205
Figure 11.10
Improvement in mean performance through training at eight di¤erent positions in the visual field for eight observers. (Left)
Mean improvement at each of the visual field positions; (right) decrease of performance at transition from one position to the
next. (After Fahle, Edelman, and Poggio 1995.) FP, fixation point. Vertical bars indicate standard errors of means.
circle around the fovea with a radius of 10 . The se- The next question was whether the specificity of
quence of testing these positions was counterbalanced perceptual learning would also be true for the eye
between observers. Within the one-hour sessions used during training. To answer this question, we
spent with each visual field position, observers im- tested another group of twelve naive observers re-
proved their performance on average by about 6–7% garding eye specificity on a vernier discrimination
(upward columns in figure 11.10). When, however, task: six observers started with the left eye patched;
they moved their gaze to the new position, there the remaining six, with the right eye patched. Per-
was a significant fallo¤ in performance (downward formance increased fastly during the first 20 minutes
columns) that matched in size the improvement of training, corresponding to the first six to eight
obtained at the previous position. Which is to say, blocks of stimulus presentations, and more slowly
observers, on average, started learning completely thereafter (figure 11.11). The last block of presenta-
anew at each visual field position. In contrast to the tions to the eye trained first, taking place usually
results on orientation specificity, there was no over- after one night’s rest, yielded the best performance,
shoot, that is, performance deteriorated, but just to indicated by the last data point before the vertical
the baseline, not below it. Thus we can safely con- line that symbolizes transition between presentations
clude that perceptual learning is specific for visual to the two eyes. After the transition, performance
field position, at least if the positions are more than fell o¤ sharply, but the rebound was faster than in
5 apart (Fahle, Edelman, and Poggio 1995; cf., how- the case of orientation change. The data indicate a
ever, Beard, Levi, and Reich, 1995). certain specificity for the eye used during training,
206 Manfred Fahle
Figure 11.11
Monocular training of a vernier discrimination task in twelve observers leads to improvement in mean performance followed
by a moderate drop of performance when the other eye is used for testing (vertical dividing line corresponds to time of
transition between eyes). (After Fahle, Edelman, and Poggio 1995.) RE, right eye; LE, left eye. Vertical bars indicate stan-
dard errors of means.
but are not completely convincing in this respect, was about as good as it was in the first eye at the end
because the fallo¤ in performance is pronounced for of training (cf. chapter 10.3.4).
only one datapoint (cf., however, chapters 5.3, 9.5,
14.3.1).
11.7 Possible Cortical Localization of the
We therefore repeated this experiment with a new
Neuronal Changes Underlying Perceptual
group of six observers, measuring thresholds during
Learning
five hours of training. Initial performance in this
group of observers was around 16 arc sec, improving
Thus far, we found that the improvement in vernier
to around 10 arc sec within the training period. At
discrimination achieved through training is highly
the transition from testing one eye to the other,
significant for the orientation of the stimulus, for its
thresholds increased sharply, again to values above
position in the visual field, and for the eye used
the initial level. This increase was again less pro-
during training. This combined specificity allows us
nounced than in the case of changing orientation,
to hypothesize about a possible localization, in the
but nevertheless highly significant (figure 11.12).
visual system, of the neuronal changes underlying
Obviously, observers had to relearn the task of ver-
the improvement in perceptual discrimination.
nier discrimination when the trained eye was cov-
ered and they had to perform the discrimination Hubel and Wiesel have shown (1959) that the ma-
jority of neurons in the primary visual cortex (area
with the eye covered during training. After another
17, or V1) of the cat and monkey have elongated
five hours of training, performance in the second eye
Learning to Perceive below Foveal Photoreceptor Spacing 207
Figure 11.12
Monocular thresholds for vernier discrimination improve over a five-hour period in six observers. Continuation of testing
with the other eye leads to an overshooting increase of thresholds, followed by a gradual relearning. RE, right eye; LE, left
eye. Vertical bars indicate standard errors of means.
receptive fields of the type displayed in figure 11.7. late body are even less probable candidates for adult
On the other hand, ganglion cells in the retina and plasticity than the primary visual cortex is (cf. chap-
neurons in the lateral geniculate nucleus, that is, in ter 1.6.1).
the structures peripheral to the primary visual cortex, On the other hand, the primary visual cortex
have circular symmetric receptive field character- seems to be the last structure—in the course of visual
istics. Because neurons on levels below the visual information processing—where information coming
cortex cannot discriminate between di¤erent ori- from the two eyes is separated, namely, within the
entations, given the somewhat unstable visual fixa- so-called ocular dominance columns, first described
tion of humans, it would be di‰cult to realize an by Hubel and Wiesel (1968). Though binocularly
improvement of performance specific for stimulus activated in all extrastriate areas, many cortical neu-
orientation by modifying these neurons. Because rons are monocularly activated in the primary visual
they would be activated by all possible stimulus ori- cortex. Thus the eye specificity of vernier learning
entations, changing the properties of neurons coding indicates that the learning takes place in the primary
for one stimulus orientation would tend to influence visual cortex, which seems able to improve function
the processing of all other stimuli projected on these through training even in adults, contrary to the view
neurons. And because we certainly would expect of a ‘‘hardwired’’ primary cortex cited above (cf.
use-dependent long-term changes in the primary chapter 10.3.4, but see Mollon and Danilova 1996
visual cortex before postulating them for the human for another view). This hypothesis of adult cortical
retina, ganglion cells and cells in the lateral genicu- plasticity is in agreement with recent electrophysio-
208 Manfred Fahle
Figure 11.13
Improvement of performance ( percent correct responses) through training in a three-dot bisection and a three-dot vernier
task. Six observers started with the bisection task; six others with the vernier task. Average percentages of correct responses
improved for both groups of observers in the first task, but there was no transfer of improvement to the second task (results
right of vertical dividing line). Rightmost data point shows results for retest of the first task. The insets show the di¤erences
between stimuli of the two tasks. (After Fahle and Morgan 1996.) Standard errors (vertical bars) were somewhat larger for
the second task, indicating larger interindividual di¤erences. Actual displacements of the central dots were far smaller than
shown here.
logical (chapters 2–4, 6–7) and anatomical results relearned for a changed stimulus orientation. To test
(chapter 1). this possibility, we performed an experiment with
two extremely similar stimuli: a three-dot vernier
task and a three-dot bisection task (see inset of figure
11.8 Possible Motor Artifacts in Perceptual 11.13). Six new observers performed a three-dot
Learning vernier task, indicating whether the middle of three
points was o¤set to the right or to the left relative to
But is all this improvement really due to sensory an imaginary line through the endpoints. The other
learning, rather than to improvement of some motor six observers performed a three-dot bisection task,
component? One could argue that observers learn, indicating whether the middle point (which was
during the course of the experiment, to improve perfectly aligned on the imaginary line between the
fixation or accommodation of the stimuli on the endpoints) was closer to the upper or to the lower
monitor, or both. According to this argument, the dot.
fallo¤ in performance observed after the change of Thresholds in these tasks were usually around 20
stimulus would be caused by a loss of applicability arc sec, corresponding to half a foveal photoreceptor
for a motor ability learned during the first part of the diameter. In the vernier tasks, the middle dot was
experiment, for example, fixation would have to be o¤set by half a receptor diameter to the right or left,
Learning to Perceive below Foveal Photoreceptor Spacing 209
whereas in the bisection task, it was o¤set by half a not trained than available at the start of the experi-
photoreceptor diameter toward the upper or lower ment. But such a mechanism, if implemented, would
dot. The two bisection stimuli di¤ered from each cause a sharp fallo¤ in performance when the first
other by more than they di¤ered from either of the task or orientation was retested after the second one
two vernier stimuli (see inset of figure 11.13). had been learned. Because this is not what we found
Moreover, the vernier stimulus o¤set to one side was when we tested di¤erent tasks or di¤erent orienta-
more similar to both bisection stimuli (regarding tions, another neuronal mechanism must underlie
Euclidean distance) than to the opposite vernier perceptual learning.
stimulus. Thus the physical stimuli in the two types
of tasks were physically very similar indeed, and any
11.9 Perceptual Learning in Amnesic
motor adaptation to one type of task should certainly
Patients
also apply to the other type of task. Observers’ per-
formance showed highly significant improvement
If perceptual learning at least partly relies on changes
during the one-hour training, falling o¤ slightly for
at an early level of visual information processing,
the retest of the old task on the second day. The
without much cognitive involvement, then even
transition to the new task was marked by a sharp
amnesic patients might be able to improve perfor-
fallo¤ in performance, once again below baseline
mance through perceptual learning. To test this
level, which gradually improved afterward through
possibility, we examined six patients su¤ering from
training on the new task. At the end of the experi-
organic amnesia due either to Korsako¤ syndrome or
ment, we retested observers on the first task. The
to ischemic damage to the anterior thalamus (Fahle
result is indicated by the rightmost data point in
and Daum in press). Because it was far more di‰-
figure 11.13, showing that performance did not fall
cult to explain the task to these observers, only five
o¤ when returning to the first task.
blocks of eighty vernier presentations each could
This experiment has two important implications:
be measured within each one-hour session. And
(1) improvement of vernier detection is indeed based
because the two sessions were separated by a gap of
on sensory rather than motor factors; (2) training on
one week, the patients had completely forgotten
the second task does not interfere with performance
about the setup, the task, and the experimenter after
on the first, ruling out a possible neuronal mecha-
the one week interval. Nevertheless, three of the
nism to explain perceptual learning. One may have
patients showed clear improvement in performance
supposed that, through training, a pool of neurons is
through training, two of them even highly signifi-
recruited for the task trained, at the expense of other
cant improvement (figure 11.14). Improvement was
tasks or other preferred stimulus orientations. This
less clear in one additional patient, whereas there was
hypothesis would predict that by training vertical
no improvement in the remaining two patients.
verniers, more cells would be devoted to analyzing
Although it comes as no suprise that the interob-
these vertical stimuli, at the expense of the cell pool
server di¤erences in these patients were even larger
analyzing horizontal stimuli. Such a mechanism
than the already high variation in normal observers, a
would explain the undershoot of performance at the
clear message of this experiment is that even patients
time of change in task or orientation. There would
su¤ering from severe amnesia, with no recollection
be fewer neurons available for the task or orientation
of having performed the task previously, can strongly
210 Manfred Fahle
Figure 11.14
Vernier discrimination thresholds as a function of training in six patients (R. E., J. R., A. S., H. W., H. S., M. H.) su¤ering
from amnesic syndromes. At least some of these patients (R. E., M. H., J. R., H. W.) improved performance. (After Fahle
and Daum in press.)
improve their vernier discrimination through train- the observers, we asked them to silently count the
ing of the task. number of presentations with verniers o¤set to the
left during part of the experiments. These o¤sets
represented a clear minority, with the large majority
11.10 A Sum-Potential Correlate of of presentations displaying o¤sets to the right. The
Perceptual Learning potentials were averaged over six hundred pre-
sentations and subsequently analyzed.
The hypothesis that perceptual learning involves The latency of the main component of the evoked
relatively early parts of the visual cortex also suggests response in ten observers decreased by around 10
that the potentials evoked over the occipital cortex msec, from 118 to 105 msec over the course of the
might change as a result of perceptual learning. To learning. Subsequently, the stimuli were rotated by
test this possibility in an electrophysiological experi- 90 . Latencies for this orientation were around 125
ment, we presented five verniers, each consisting of
msec after the transition, decreasing to around 114
three elements that were first perfectly aligned, but
msec over half an hour of training. The combined
whose middle segments were displaced slightly to
e¤ect was significant on the 5% level.
one side at a defined time. The appearance of the
We statistically compared the distributions of
o¤sets evoked cortical potentials especially over the
potentials over the occipital pole for the first versus
occipital pole of the brain; we recorded these evoked
second 600 presentations of the stimulus. The dif-
potentials by an array of sixteen electrodes (see Fahle
ferences between pre- and posttraining distributions
and Skrandies 1994). To increase the attention of
Learning to Perceive below Foveal Photoreceptor Spacing 211
Figure 11.16
Orientation bandwith of perceptual learning. Change in stimulus orientation for di¤erent groups of observers, from 90
down to 2 , plotted along abscissa; resulting amount of transfer of improvement (squares) and improvement obtained with
the first stimulus orientation (circles), plotted along ordinate.
learning by plotting the di¤erences in start levels for man observers yielded similar values (e.g., Campbell
the first versus second stimulus orientations, plus the and Kulikowski 1966; Movshon and Blakemore
improvement of performance during training with 1973). Because the orientation specificity of percep-
the first orientation. Complete transfer has been tual learning is thus far more pronounced than that
achieved only if the di¤erence between the start levels of single neurons or neuronal populations as mea-
equals the improvement obtained at the first level. sured by adaptation techniques, it is probably not the
These findings indicate that orientation specificity receptive field of primary visual cortex neurons that
of perceptual learning is on the order of less than is modified as a result of perceptual learning, and we
10 , probably around 5 . Which is to say, the neu- have to dismiss the speculative explanation put for-
ronal mechanisms underlying perceptual learning ward in 11.4.
must have a half-width that is even lower. This high Another argument against the hypothesis of adapt-
specificity for orientation contrasts quite sharply with ing early filters comes from a study investigating
the orientation specificities of single neurons in the transfer between three hyperacuity tasks, namely
primary visual cortex of, for example, monkeys. De vernier discrimination, orientation discrimination,
Valois, Yund, and Hepler (1982) found the half- and curvature detection. Eighteen observers were
width of single neurons to have a median around 20 sequentially trained on each of these three hyper-
(with a few cells showing bandwidths of only 3 acuity tasks for one hour in counterbalanced order.
half-width). Adaptation and masking studies in hu- The results, as shown in figure 11.17, panel b, show
Learning to Perceive below Foveal Photoreceptor Spacing 213
Figure 11.17
Training of three hyperacuity tasks: vernier (left), orientation (middle), and curvature (right) discrimination. The six groups of
observers trained these tasks in counterbalanced order, without indication of transfer between the tasks (at the times indi-
cated by vertical dividing lines). (a) Orientation cues inherent in the three types of stimuli; (b) mean performances and
standard errors of all observers for the three tasks in chronological order. (After Fahle 1997.)
a clear improvement of performance during the first gradually during the one-hour training phase of the
session. In each of the sessions, the same number of second task. Basically, the same was true for the
observers were trained on each of the three hyper- transition to the third task. Perceptual learning thus
acuity tasks. The first measurement on the second displays a high specifity for the exact type of hyper-
day of testing (nearest point to the right of the acuity task. If, on the other hand, perceptual learn-
dotted vertical line) proved that there was no change ing sharpened the tuning characteristics of the early
of performance during the night. When subse- orientation selective filters, such as simple cells in the
quently the next hyperacuity task was tested, per- primary visual cortex, then, because all three hyper-
formance fell o¤ to pretraining levels, improving acuity tasks can be solved by using orientation in-
214 Manfred Fahle
formation, we would have expected some form of possible even without error feedback, these experi-
transfer of improvement (see figure 11.17, panel a). ments employed stimuli well above discrimination
threshold. Thus some form of backpropagation may
have taken place, that is, correct perception of some
11.12 Perceptual Learning Is Based Partly on
of the clearly identifiable stimuli may have been used
Top-Down Influences: Feedback
to identify less clear presentations. We therefore
performed an experiment with vernier o¤sets of only
We have seen that a simple, straightforward assump-
10 arc sec, below the perceptual thresholds of most
tion regarding the neuronal mechanisms underlying
observers. As can be seen from figure 11.18, the ini-
perceptual improvement through learning and train-
tial level of performance was around 60% correct
ing cannot be true. Moreover, important theoretical
responses, not far above the chance level of 50%.
considerations would contradict the assumption of The six observers who received error feedback
early filter adaptation. If these early filters were in-
( panel b) improved rapidly, within less than ten
deed permanently modified as a result of training,
blocks, and slowly thereafter. But observers not
then these modifications would influence the future
receiving error feedback were also able to improve
processing of all previously learned visual patterns,
performance, even if more slowly than the ones
which in turn might be detrimental for the detec-
receiving feedback ( panel a) (cf. chapter 13.3). At
tion of those earlier patterns. To avoid massive cross-
the end of the one-hour training period, both groups
interference between the processing of di¤erent
of observers yielded about the same level of correct
patterns, the visual system should keep the early
responses. Thus it seems that perceptual learning is
filters constant, as was already clearly stated by Marr indeed possible without feedback, even close to
(1982). How, then, can perceptual learning be so
threshold, but that feedback increases the speed of
highly stimulus specific? How can it indicate a neu- learning.
ronal site at a relatively early level of visual infor-
How might feedback signals achieve this speeding
mation processing—being specific even for the eye
up of perceptual learning? One straightforward hy-
used during training—yet avoid modifying the early
pothesis is that feedback acts as a teacher, helping in
orientation-specific filtering mechanisms?
the classification of each vernier as o¤set to the right
The only resolution to this dilemma, in my opin-
or to the left by providing observers all the infor-
ion, is to suppose the influence of strong feedback
mation they need to correctly classify—after the
signals stemming from higher levels of visual infor-
response—the vernier’s o¤set direction. In the next
mation processing and acting on the primary visual
experiment, we therefore omitted half the error sig-
cortex (see Herzog and Fahle 1997 and chapter 20, nals, that is, half the stimulus presentations incor-
this volume). We performed a number of experi-
rectly identified by the observer did not receive error
ments to determine whether higher cortical levels feedback. Thus these stimuli would not only be
actually influence perceptual learning. The first class incorrectly classified at first, but this incorrect classi-
of experiments concerned error feedback. Although fication would be confirmed because the lack of
it had been shown previously (McKee and West- error feedback would indicate a correct classifica-
heimer 1978; Fendick and Westheimer 1983) that tion. Any learning strategy based on a teacher signal
improvement through training in hyperacuity tasks is would su¤er from severe misclassifications of the
Learning to Perceive below Foveal Photoreceptor Spacing 215
Figure 11.18
E¤ect of training with subthreshold vernier stimuli, plotted as mean performances of six observers. (a) Without error feed-
back; (b) with error feedback. Performance improves significantly under both conditions. (After Fahle, Edelman, and Poggio
1995.) Vertical bars indicate standard errors of means.
11.14 Conclusions
Figure 12.2
Schematic illustration of the correlation between receptive
field size and the anatomical level of the visual hierarchy
(not drawn to scale). With each level of processing, the
mean receptive field size (ellipses) of neurons becomes
larger. V1, V2, V4, PIT ( posterior inferotemporal) cortex,
and AIT (anterior inferotemporal) cortex are areas within
the visual pathways of the brain. Many cells in higher areas
respond to the same stimulus in ways largely independent
of its location within the visual field.
Figure 12.1
Schematic drawing of the central portion of the visual
field. Vision is sharpest in the center of the fovea (sym- visual cortex (area V1), the major input stage for vi-
bolized by the black circle), but acuity progressively drops sual information to the brain, have relatively small
o¤ toward the periphery (gray and white concentric rings). receptive field centers of about 1 diameter—at least
The distance from the fovea is measured in degrees of
visual angle.
in the central parts of the visual field. Cells in areas
V2 and V4 have progressively larger receptive fields,
normally subtending over several degrees of the
in the visual brain. A functional consequence of this visual field. Finally, cells in anterior inferotemporal
organization is that, depending on the position of a (AIT) cortex often respond to stimuli within 20 or
stimulus in the visual field, the image of the stimulus is more. Indeed, AIT cells can cover nearly the whole
projected on di¤erent parts of the retina and stim- visual field; the anterior inferotemporal cortex is
ulates di¤erent cell populations in primary visual generally considered to be responsible for high-level
cortex (area V1) and higher brain areas. pattern and object recognition.
It is a remarkable ability of the human brain that it
can—at least in principle—identify objects regardless
of where they are presented in the visual field. Ap- 12.2 The Rationale of Position Transfer
parently, the brain achieves position invariance of Experiments
recognition in several steps. With each level of visual
processing, the mean size of receptive fields becomes The correlation of receptive field size and level
larger and larger (figure 12.2). Cells in the primary of processing o¤ers an elegant way to investigate
Specificity versus Invariance of Perceptual Learning 221
12.3.1 Overview
Table 12.1
Summary of results of various position transfer experiments
Smallest
position Transfer
Learning paradigm References shift tested (F/P/N) Notes
Stereopsis Ramachandran 1976 2.5 N/P Some subjects no, others partial
transfer.
Stereopsis O’Toole and Kersten 1992 0.7 N (P?) Give no baseline; influence of
selective spatial attention?
Popout detection Ahissar and Hochstein 1996, 0.7 N/P Transfer to homologous positions
1997 across the midline; correlation of
positional specificity and task
di‰culty.
Vernier discrimination Fahle 1994; Fahle, Edelman, 10 N Only with trained observers;
and Poggio 1995 unspecific learning with naive
observers.
Vernier discrimination Beard, Levi, and Reich 1995 2 P Some transfer along meridians;
due to spatial attention?
Texture discrimination Karni and Sagi 1991 3 N
Pattern discrimination Nazir and O’Regan 1990; 1 N
Dill and Fahle 1997
Line orientation Shiu and Pashler 1992 11 N Cognitive influences.
discrimination
Grating orientation Schoups, Vogels, and Orban 2.5 N (P)
discrimination 1995
Classification of Gabor Rentschler, Jüttner, and 3 P Some transfer to homologous
patches Caelli 1994; Jüttner and positions across the midline.
Rentschler 1996
Motion discrimination Ball and Sekuler 1987 2 P Some transfer even without
stimulus overlap.
Spatial phase discrimination Berardi and Fiorentini 1987 1 N/P Transfer to homologous positions
across the midline.
224 Marcus Dill
Figure 12.6
Positional specificity in vernier (left) and pattern discrimi-
nation (right) learning. White bars represent the improve-
ment of performance due to perceptual learning. Black
bars show the results directly after the position shift com-
pared to base level.
Figure 12.7
Example of target pattern (top) and distractor set (bot-
dots. When one of these stimuli is presented to a tom). Each pattern is bilaterally symmetric and consists of
defined position in the visual field and subjects have a 6 6 matrix of elements that are filled with a large black
to decide whether a specific pattern—the target rectangle at 50% probability. One random pattern is
pattern—or one of several distractor patterns has defined as the target; eighteen distractors are created by
changing a single element from black to white or vice versa.
appeared, their responses are highly erroneous at the
beginning of the experiment. Subjects thus have
di‰culty discriminating novel dot patterns, but can grating and vernier learning (see also section 12.5,
learn to do so in training experiments. where pattern discrimination learning is discussed in
In a typical experiment (Nazir and O’Regan 1990; greater detail and these experiments related to how
Dill and Fahle 1997), performance improves within a invariance of pattern recognition to retinal trans-
few hundred trials, depending on the complexity of lations is achieved).
the stimuli and on other experimental conditions.
When subjects eventually reach a criterion level of, 12.3.3 Exceptions
say, 90% correct responses, training stops and they
are tested at either the training or a transfer location. Task Di‰culty
As shown in figure 12.6, performance at the transfer Although most studies found no transfer or only
location falls o¤ significantly compared to the train- partial transfer to new locations, there are a few
ing location, with accuracy at the new location no noteworthy exceptions. For example, Ahissar and
better than the base level at the beginning of the Hochstein (1997) showed that task di‰culty can be
training. Pattern discrimination learning is thus com- one factor influencing the degree to which learning
pletely specific to the location of training, as was is position specific. Simply stated, learning an easy
226 Marcus Dill
task transfers to new locations, whereas learning a Spatial attention can thus lead to positional specif-
more di‰cult task does not. Obviously, lower levels icity in both behavioral and physiological perfor-
of processing get involved as soon as the task is too mance. Could it be that by training the visual system
di‰cult to be solved completely by higher levels. It at just one location it is pushed to attend to this area,
has been suggested (Ahissar and Hochstein 1997) while ignoring the remaining parts of the visual
that learning follows the visual processing hierarchy field? If this is the case, positional specificity would
in a reverse direction: first, units on higher levels have to be considered as an attentional artifact rather
are selected that best solve the experimental task; than as a characteristic of visual learning mechanisms.
then, if these units are not su‰cient to achieve reli- O’Toole and Kersten (1992) investigated this
able performance, they recruit additional resources possibility with regard to the positional specificity
from lower levels. However, the trade-o¤ for fur- they found for learning stereoscopic depth percep-
ther improvement of performance from the partici- tion. In their original experiments, they trained
pation of extra capacities is a lack of transfer to new subjects at a single location and found reduced per-
locations. formance after transfer to another location. In a
control experiment, however, they trained subjects
Selective Spatial Attention with two patterns at two di¤erent locations. For
A second factor assumed to influence the degree of example, pattern A was used only at position 1,
positional specificity is selective spatial attention: fo- whereas, within the same training session, pattern B
cusing attention on a particular location in the visual was trained exclusively at position 2. Later, both
field can improve performance in visual tasks within patterns were recognized equally well at both loca-
a delimited window of attention while diminishing tions. These additional results show that the acquired
perceptual abilities in other areas of visual space. This ability to discriminate depths can transfer as long as
well-established phenomenon, selective spatial at- enough resources are allocated to both locations.
tention, has been described in many classical studies. Thus positional specificity in this class of experi-
The ability to detect a stimulus increases significantly ments can, indeed, be attributed to selective spatial
when the location of this stimulus is indicated be- attention rather than to depth perception.
forehand by a position cue preceding the stim- That vernier learning partially transfers along
ulus. On the other hand, in trials with an invalid meridians of the visual field (Beard, Levi, and Reich
position cue—one that misleads attention to a (1995) indicates that the window of attention extends
wrong location—subjects’ ability to detect the stim- away from the fovea in a clublike fashion. Because,
ulus is even lower than with no or neutral cues. however, the extent of the attentional spotlight has
Interestingly, Moran and Desimone (1985)—and not been assessed independently in this study, it is
since then many others—found electrophysiological di‰cult to decide whether the partial transfer is the
correlates for this behavioral phenomenon. Many consequence of attentional mechanisms or of spe-
cells in V4 and other visual areas alter their response cializations of visual perception.
behavior in a similar manner. If attention is directed If attention plays a role in vernier learning at all,
to one visual field location, stimulation of other this role may be more general in the sense that ob-
locations may be without e¤ect, although the stim- servers learn how to attend to stimulation outside of
ulus still lies in the classical receptive field of this cell. the center of the fovea. Fahle and coworkers (Fahle
Specificity versus Invariance of Perceptual Learning 227
1994; Fahle, Edelman, and Poggio 1995), for exam- and Fiorentini 1987; Rentschner, Jüttler, and Caelli
ple, reported that subjects who had never before 1994; Ahissar and Hochstein 1996) found some
taken part in psychophysical experiments involving transfer or even complete transfer of learning to ho-
extrafoveal stimulation showed some transfer of mologous locations across the vertical midline of the
learning to other peripheral or parafoveal locations. visual field. This special role of symmetrical positions
This unspecific e¤ect can most probably be attrib- seems not to be attributable to callosal connec-
uted to learning how to divert attention away from tions between the two hemispheres. Fiorentini et al.
the center of the visual field, where it is normally (1992) reported the case of a boy born without a
focused. Once subjects possess this know-how, fur- corpus callosum who showed the same pattern of
ther improvement of performance remains specific transfer as normal subjects: complete transfer to sym-
to the location of training. metrical locations, but no transfer to other places
For most of the perceptual learning paradigms, (cf., however, chapters 9, 10). Besides, a special role
there are good arguments against an artifactual in- of positional symmetry relative to the fovea has also
fluence of selective attention. As mentioned above, been found for the vertical dimension, that is, for
some studies tested transfer over less than 1 dis- symmetrical positions above and below the center of
placement and found complete positional specificity. the fovea (Dill and Fahle 1999). Although the latter
Experiments testing the minimal size of the window experiments did not involve learning, they suggest
of spatial attention, on the other hand, have shown that the above findings are not the consequence of a
that observers cannot split attention between stimuli special interhemispheric transfer.
that lie within 1 of the visual field. Thus, at least in
some cases, selective spatial attention can be ex-
12.4 What Positional Specificity Tells Us
cluded. Besides, there is no electrophysiological evi-
dence to date that the modification of the receptive about the Locus of Plasticity—and What It
Does Not
field by spatial attention is permanent. One might
also expect that this tuning of high-level elements
According to the rationale outlined in section 12.2,
would deprive other spatial positions of processing
most authors have concluded that perceptual learn-
resources, leading to performance levels even below
ing involves not only higher levels of the visual hi-
the starting level after change of the location. There
erarchy, but also lower levels. Combining positional
is no evidence, however, for an undershooting e¤ect
specificity with other results such as the orientation
after spatial displacements: performance after posi-
specificity and—in some cases—the monocularity of
tion shift was never worse than at the beginning of
training. Most studies seem to prove that learning perceptual learning, some have argued that even the
primary visual cortex, the very first visual area of the
at di¤erent locations is independent even in this
brain, may show some plasticity.
respect.
Although these interpretations may well be justi-
fied, when speculating on the locus of learning, one
Symmetrical Locations
also has to be aware of several critical issues related to
Finally, a positional ‘‘anomaly’’ for symmetrical loca-
these tests for positional specificity. For example,
tions is worth noting. At least three studies (Berardi
a number of experiments tested transfer only very
228 Marcus Dill
defined by presenting it ten times to the subjects. ing process is necessary. During training, subjects
Later, the target is shown in half of the trials and a refine their internal representation of the patterns
randomly selected distractor pattern in the other step by step: they notice more and more discrim-
half. Each distractor di¤ers from the target by only inative features of a target (or of distractors), so that
one element and its mirror-symmetric counterpart, their performance improves until, after some time,
which are changed from black to white or vice versa they consistently give correct responses. Even if feed-
(see figure 12.7). back is no longer given, accuracy stays at high levels as
Because the presentation time (@150 msec) is too long as the position of presentation is kept constant.
short for scrutinizing the patterns by directed eye As we have noted, when pattern location is
movements, subjects at first do not see much more changed, performance falls back, not to chance, but
than an undefined cloud of dots. In most trials at the to the base level subjects reached at the beginning of
start of the training, observers can only guess, al- the training. As at the beginning of the training,
though, in some cases, they may recognize ‘‘some- subjects after the position shift do not simply guess.
thing’’ in a particular pattern—much in the same Some information regarding the particular stimulus
way that we perceive, say, a face or an animal in the must have been transferred to the new location. That
clouds. The recognized content may be an associa- performance after transfer is remarkably similar to
tion with a familiar object, a simple shape, a charac- initial performance makes it very likely that III
teristic elongated curve, or a particular configuration is position invariant. Any discrimination ability ac-
of dots. Many of these random images resemble quired during the course of training, on the other
human or animal faces, most probably due to their hand, is apparently specific to the location where it
mirror symmetry. If the target looks like a strange has been learned.
monkey face, say, it will be very easy to distinguish it But why should the immediate input interpreta-
from a distractor whose appearance is not facelike tion be position invariant, whereas perceptual learn-
because the ‘‘eyes’’ (two symmetric dots), present in ing is not? Introspection and common sense support
the target, are now absent. In other cases, an isolated the conclusion that III is position invariant. Because
black dot surrounded by white fields may be very both the facelike appearance of a pattern and the
prominent in the target, so that it is rather easy to existence of an isolated dot or another striking fea-
note when it is absent in a distractor. ture within it are conscious perceptions that can be
This spontaneous recognition of familiar objects communicated by subjects, it is reasonable to assume
or simple shapes in an apparently novel pattern—let these high-level or even abstract characteristics of
us call it ‘‘immediate input interpretation’’ (III)— dot stimuli are detected independently of their loca-
helps subjects to identify at least some features, tion within the visual field. Subjects do not have to
though not the whole stimulus domain, already recognize that specific face or isolated dot again.
during the ten presentation trials before the actual They need only detect ‘‘a’’ face or ‘‘a’’ single dot.
training. Quite often, therefore, initial performance That perceptual learning is position specific can be
is not at chance level, not even during the very first rationalized by the reverse hierarchy model (Ahissar
trials. III may allow reasonable, although not always and Hochstein 1997) mentioned above and ex-
perfect perceptual performance. To achieve a higher plained in detail in chapter 14 of this volume. Dur-
level of accuracy in our discrimination task, a learn- ing the first trials of an experiment, the visual system
Specificity versus Invariance of Perceptual Learning 231
the other hand, this point of view places the mech- 13.1a, readers may look at the original gray-scale
anisms underlying perceptual learning at a level quite image, in figure 13.1b, as a visual hint, and then go
separate from those that are traditionally associated back to figure 13.1a to experience the changed per-
with insight. The sudden improvements in perfor- ceptual organization.) Using a task that involves a
mance observed in insight phenomena have been similar ( but much-simplified) transition in perceptual
taken to indicate a cognitive event that occurs more organization, we will show that, under appropriate
centrally. Insight seems to be a process involving the experimental conditions, it is possible to cause the
whole animal, as in Köhler’s apes (1925), who sud- ‘‘sharp drops’’ in error curves mentioned by Hebb to
denly realize the potential relation of disparate ele- occur at a predictable time. The ability to gain ex-
ments in their visual field. perimental control over when the unique event of
Nevertheless, the two forms of learning may be insight occurs addresses a long-standing problem that
more related than is currently thought. Hebb (1949) researchers have faced, and suggests that perceptual
suggested that the dichotomy between insight and learning may be a particularly suitable paradigm to
rote learning may be artificial, and that the two study the role of insight in learning.
forms of learning may share common mechanisms.1 We will also address the relation between stimulus
Hebb asked: ‘‘Is insight or hypothesis—or, in the specificity and insight in perceptual learning. As dis-
broadest terms, intelligence—something distinct cussed in chapters 9 and 10 of this volume, it is often
from the mechanism of association?’’ ( p. 163). He found in studies of perceptual learning that the im-
observed that ‘‘learning is often discontinuous; error provement in performance does not generalize across
curves show sharp drops without warning, and the stimulus attributes such as retinal size and location,
kind of error that is made on one day may be quite or shape (Ramachandran and Braddick 1973; Fior-
changed in the next’’ ( p. 159). He concluded that entini and Berardi 1980; Karni and Sagi 1991; cf.
‘‘insight . . . continually a¤ects the learning of the chapters 9–12). Indeed, these findings have played a
adult animal’’ ( p. 163), and that ‘‘it is not wholly major role in driving theories that place the site of
separate from rote learning’’ ( p. 164). Are Hebb’s plasticity in perceptual learning at early cortical areas
assertions valid for the case of perceptual learning? Is (e.g., area V1; Fiorentini and Berardi 1980; Karni
there evidence that insightlike phenomena are part and Sagi 1991). At first glance, we might expect
of the process of learning to perform a perceptual improvements from insight not to be susceptible to
task? In this chapter, we will argue that the answer to such superficial changes of circumstances and thus to
these questions is yes. Although we are used to think generalize to stimuli that di¤er only in low-level
of insight in the context of high-level cognitive tasks visual properties. But in our experiments, the in-
(such as problem solving), abrupt improvements in sightlike abrupt learning does not generalize to a
performance, resembling an occurrence of insight, new retinal size. This finding echoes reports from
can be observed in visual perception, as well. A the literature on problem solving, that subjects’
classic example is the perception of hard-to-segment ability to generalize an insightful solution, that is, to
pictures, such as the one shown in figure 13.1a, transfer the solution to a novel context, depends on
where a dramatic transition to the ‘‘correct’’ inter- the extent of surface-level similarity between prob-
pretation may occur spontaneously or as a result of a lems (Gick and Holyoak 1980; see also Ippolito and
cognitive or visual hint. (After looking at figure Tweney 1995 on the specificity of expert insights).
The Role of Insight in Perceptual Learning 237
Figure 13.1b
Original gray-level image from which figure 13.1a was pro- Figure 13.2
duced can be used as a ‘‘visual hint’’ to see the embedded Shape discrimination task based on the perception of illu-
figure. sory contours. The inducers of a Kanizsa square were
rotated about their centers by a variable degree, resulting
in the perception of curved illusory surfaces of ‘‘fat’’ (left)
as in the experimental blocks). The practice session or ‘‘thin’’ (right) shapes. Observers were required to choose
consisted of four examples of long-duration stimuli, between the two alternatives. (a) Direction and degree of
followed by 20 presentations of brief, masked stim- inducer rotation determined the sign and amount of cur-
vature of the illusory surfaces, respectively (our convention
uli. Subjects were required to give at least 17 out of
is to denote the direction of rotation that produced ‘‘fat’’
20 correct responses in their first or second practice surfaces as negative). The range of curvatures used was
block in order to participate in the experiment (3 varied from one experimental block to the other, thus
out of subjects 33 were rejected from the experiment allowing for control of the level of di‰culty and the onset
due to failure on this criterion). Once they passed of the abrupt learning. (b) Each trial consisted of a brief
this criterion, they were given the experimental stimulus presentation of either a ‘‘thin’’ or ‘‘fat’’ surface
from the range of curvatures used in that block, followed
blocks, where the IC stimuli were less salient, be-
by a blank screen and then a mask, which was designed to
cause of the smaller inducers’ size. Subjects were interfere with the perception of the inducers but not the
given feedback in the form of a computer beep after global illusory surface. The stimulus was presented for
correct responses throughout the practice and all 97 msec in the first experiment reported here, and for 97–
experimental blocks. 194 msec in subsequent experiments (see text). The blank
screen and the mask were presented for 69 msec and
Figure 13.3 shows the performance of an individ-
250 msec, respectively.
ual observer (A. H.) in seven consecutive blocks. The
probability that the subject judged a given stimulus
(i.e., a given value of inducers’ rotation) as ‘‘thin’’
was computed for the twelve repetitions of that
The Role of Insight in Perceptual Learning 239
mixed. As is evident from panel b, the observer’s unchanged, the greater viewing distance meant that
performance improved dramatically: the threshold in the retinal size of the stimuli became smaller: the side
this block was 1.8 (the data from longer exposure of the illusory surfaces was now 5.7 visual angle
duration stimuli were not included for the calcula- (inducers’ size: 1.4 ), compared to 14.3 visual angle
tion of the threshold). That the subject performed (inducers: 3.6 ) at the 60 cm viewing distance. The
well on the higher-curvature stimuli is to be ex- first block at this new viewing distance (‘‘test’’; panel
pected because they are inherently easier to discrim- e) consisted of low-curvature stimuli, like those in
inate. But note that performance improved markedly the ‘‘test’’ and ‘‘retest’’ blocks at the 60 cm viewing
also on the low-curvature stimuli: A. H. correctly distance. Performance fell markedly compared to
discriminated in this block between ‘‘thin’’ and ‘‘fat’’ before (threshold: 5 ). Thus the learning observed in
figures with curvature values of 2 and 3 in 92% and the 60 cm viewing distance did not generalize to the
96% of the trials, respectively, compared to only 58% new retinal size. To ensure that good performance
and 63% on identical stimuli in the previous block. was in fact possible for this smaller retinal size, the
This dramatic improvement was not due to a lack subject was given a ‘‘train’’ block similar to that used
of cognitive understanding of the task in the first at the 60 cm viewing distance, where high-curvature
block—A. H. got 20 out of 20 trials correct in the (4–6 ) stimuli were mixed in with the low-curvature
practice block. Would good performance on low- stimuli ( panel f ). This procedure again triggered a
curvature stimuli always require the presentation of rapid improvement, leading to similar performance
high-curvature stimuli in the same block? This was to that observed before (threshold: 1.7 ). The final
tested in the third block (‘‘retest’’; panel c), which block at the 150 cm viewing distance was again a
consisted of a stimulus set identical to that of the repeat of the low-curvature stimulus set (‘‘retest’’;
first block. The good performance was maintained panel g), but this time performance was good
(threshold: 1.6 ), indicating that, compared to the (threshold: 1.8 ), indicating that the subject was able
poor performance exhibited in the ‘‘test’’ block, A. to learn the task at the new retinal size as well.
H. had undergone a rapid process of perceptual Figure 13.4 summarizes the results of six naive
learning: he could now perform well on a set of observers who were given the same sequence of
stimuli that were too di‰cult for him before. blocks as A. H., in terms of threshold performance as
The remaining four experimental blocks were run a function of block type. All subjects showed sharp
a week later. The fourth block (‘‘retain’’; figure 13.3, improvement in the transition from the ‘‘test’’ to the
panel d) was again a repeat of the low-curvature ‘‘train’’ blocks, and a lack of generalization of the
stimulus set. The good performance was maintained learned performance to the new retinal size. As dis-
(threshold: 1.5 ), indicating that the learning ob- cussed earlier, the level of performance after the
tained a week before was long-lasting. Immediately learning indicates that the subjects were basing their
following the ‘‘retain’’ block, the subject was moved judgments on perceived illusory contours, whereas
to a new viewing distance of 150 cm, and here we the poor level of performance in the ‘‘test’’ blocks is
found that the long-lasting perceptual learning characteristic of a strategy based on making judg-
described above was specific to the trained retinal ments based on the di¤erences in the local inducers’
size. Because the size of the stimuli on the screen was orientation.
The Role of Insight in Perceptual Learning 241
Figure 13.5
Trial-by-trial analysis of performance as a function of time. Percent correct discrimination for each successive pair of þ=2
(top) and þ=3 (bottom) curvature stimuli was separately tabulated, averaged over all six observers, and plotted as a function
of time. Trials 1–12 were in the ‘‘test’’ block at the 60 cm viewing distance. Trials 13–24 (shaded zone) reflect the perfor-
mance when identical stimuli were given in the ‘‘train’’ block, embedded in a set of higher-curvature and longer-duration
stimuli. Trials 25–36 were in the ‘‘retest’’ block, which consisted of a stimulus set identical to that of the initial ‘‘test’’ block.
The last four blocks, again of twelve trials each, were performed on a second session, between one and seven days later for
di¤erent observers. The ‘‘retain’’ block (trials 37–48) was performed at the 60 cm viewing distance and the next three blocks
were performed right after it, at the 150 cm viewing distance. Poor performance in the ‘‘test’’ block (trials 49–60) indicates
that the learning did not transfer to the new viewing distance, although learning subsequently reoccurred in the ‘‘train’’
(trials 61–72) and ‘‘test’’ (trials 73–84) blocks for the new retinal size.
Braddick 1973; Ramachandran 1976; Fiorentini and shows the trial-by-trial analysis of the RTs to the
Berardi 1980; Ball and Sekuler 1982; Karni and Sagi þ=2 (top panel) and þ=3 (bottom panel)
1991; Ahissar and Hochstein 1993, 1997; cf. chapters stimuli. The pattern of performance parallels that
9, 11, 12). What is the relation between the abrupt, found in the percentage correct data (figure 13.5),
insightlike learning we observed and the more gradual with a sharp drop in mean RTs at the transition be-
form of learning reported in other studies? We shall tween the ‘‘test’’ and ‘‘train’’ blocks at the 60 cm
return to the implications of this distinction in sec- viewing distance, an increase in mean RTs as sub-
tion 13.5. jects were moved to the 150 cm viewing distance,
Subjects’ reaction times (RTs) also reflect a sud- and finally again a drop in mean RTs on the ‘‘train’’
den, but stimulus-specific improvement. Figure 13.6 block at the new viewing distance. In addition, a
The Role of Insight in Perceptual Learning 243
Figure 13.6
Trial-by-trial analysis of reaction times (RTs). Mean RTs for each successive pair of þ=2 (top) and þ=3 (bottom)
curvature stimuli are plotted as a function of time (average of six observers). At the 60 cm viewing distance, trials 1–12 were
in the ‘‘test’’ block; trials 13–24 (shaded zone), in the ‘‘train’’ block, where a marked drop in the RTs is observed; trials 25–
36, in the ‘‘retest’’ block; and trials 37–48, in the ‘‘retain’’ block. At the 150 cm viewing distance, trials 49–60 were in the
‘‘test’’ block, where mean RTs increased and percentage correct performance fell o¤, as also observed in figure 13.5; trials
61–72, in the ‘‘train’’ block, where facilitation is observed again; and trials 73–84, in the ‘‘retest’’ block.
slight increase in mean RTs can be observed in the The trial-by-trial analysis reveals a course of im-
‘‘retest’’ blocks (trials 25–36 and 73–84), indicating provement that follows closely Hebb’s behavioral
that observers were aware that these were more dif- criterion (1949, p. 160) for ‘‘insight’’: ‘‘There is a
ficult than the preceding ones (‘‘train’’), although this period first of fruitless e¤ort in one direction, or
increase in di‰culty is not manifested in the per- perhaps a series of solutions. Then suddenly there is a
centage correct performance (figure 13.5). Note that complete change in the direction of e¤ort, and a
the subjects were not told that their reaction times cleancut solution of the task.’’ As mentioned earlier,
were being recorded; the only emphasis in the in- the high thresholds in the ‘‘test’’ block are charac-
structions was on the correctness of responses. In teristic of performing the task based on the local
other words, the sharp drops in the mean values and inducers’ orientation, whereas the good performance
variability of the RTs occurred even though the later indicates judgments based on illusory contour
subjects were not instructed to respond as fast as perception. This finding suggests that the improved
possible, and suggest that a facilitation in performing performance was indeed associated with a changed
the task took place. strategy, or ‘‘direction of e¤ort,’’ as Hebb suggested.
244 Nava Rubin, Ken Nakayama, and Robert Shapley
The subjective reports of the observers are consistent improvement in performance was specific to the
with this idea. Several subjects reported that, in the trained retinal size, and retraining was necessary at
first block, they did not see the global illusory the new retinal size. Thus there seems to be an in-
shapes, and were basing their judgments on the local teraction between the low-level (exposure to spe-
inducers; in the second block, they suddenly started cific stimuli) and high-level (strategy, knowledge
seeing the global shapes (sometimes noting the well- about the level of correctness) aspects of the abrupt
known brightness e¤ect associated with it; Kanizsa learning; we shall return to this point in section 13.5
1979; see also Petry and Meyer 1987). Thus both the (cf. also chapters 11, 20).
subjective reports and the behavioral measures are
consistent with a transition in subjects’ strategy in
13.4 Will Any ‘‘Easy’’ Stimulus Set Trigger
performing the task, somehow triggered by the in-
troduction of the ‘‘train’’ block. Interestingly, there Abrupt Learning?
was a notable di¤erence between subjects who were
We have seen that the abrupt learning did not gen-
practiced psychophysical observers ( but were still
eralize to a new retinal size, that is, the training
naive about the purpose of our experiment), all of
procedure was e¤ective only for the retinal size of
whom reported a transition in their strategy, com-
the stimuli used in the ‘‘train’’ block. Next we ex-
pared with unpracticed subjects, who were much
amine further the extent to which the abrupt learn-
more likely to ascribe their improvement to their
ing was sensitive to the specific attributes of the
belief that the second block was ‘‘easier.’’ That in-
stimuli in the ‘‘train’’ block. First, we asked whether
sightlike behavior can be triggered experimentally
by appropriate ‘‘hints,’’ even when subjects are un- learning would take place when the large-curvature
illusory surfaces had the same retinal size as before,
aware of the hints, has been known in the domain of
problem solving for a long time (Mayer 1995). but the inducing elements were of a di¤erent size.
To test this, we ran a new group of ten subjects,
Our results suggest that insightful learning may
which we designated ‘‘group B,’’ on the first three
not be limited to domains such as problem solving,
experimental blocks of the ‘‘thin/fat’’ task (i.e., only
but rather may play a role in perception as well. This
the first session, at 60 cm viewing distance; the sub-
view is further supported by our findings about the
jects received a practice block first, as before). The
role of external feedback in learning. During our
experiment performed by group B was identical to
pilot studies, we ran di¤erent subjects with and
the first session in the experiment described before,
without feedback, and found that, on average, sub-
except for the following change: the diameter of the
jects who did not receive external feedback about
their correctness did not show as robust learning as inducers of the high-curvature stimuli (4–6 ) was
those who did. Again, this finding was particularly increased, leaving their centers at the same locations
true of subjects who were not practiced psycho- as before, so that the support ratio was 0.4. This
physical observers; in contrast, two practiced (but change in diameter meant that the retinal size of the
naive) subjects showed an abrupt and long-lasting inducing elements was di¤erent from that used for
improvement in the absence of any external feed- the low-curvature stimuli, whereas the size of the
back. On the other hand, recall that the insightlike illusory surfaces was the same for the two types of
stimuli (see illustration in middle row of figure 13.7,
The Role of Insight in Perceptual Learning 245
Figure 13.7
Specificity of the training stimuli. (Bottom) Schematic diagram of the di¤erent procedures used for the three experimental
groups. Group A was given a ‘‘train’’ block with large-angle stimuli of the same inducer size as the small-angle (‘‘test’’) stimuli
(and those in the first three blocks in figures 13.3–13.6). Group B was given a ‘‘train’’ block where the high-curvature stimuli
were of larger-size inducers than the ‘‘test’’ stimuli. Group C was given a ‘‘train’’ block that contained long-duration low-
curvature stimuli, and no high-curvature stimuli (the longer-duration stimuli are illustrated here schematically by higher
contrast). (Top) Subjects in group A (left) and group C (right) show a dramatic improvement in their performance, which is
maintained after the training stimuli are again removed (‘‘retest’’). Subjects in group B (middle) show large individual di¤er-
ences; many do not improve during the ‘‘train’’ block at all, and those who do improve during the training block do not
retain the good performance once the large-angle stimuli are taken away (‘‘retest’’ block). The thresholds for the ‘‘train’’
block were estimated based on the data from the 1–3 short-duration stimuli only, for all three experimental groups.
246 Nava Rubin, Ken Nakayama, and Robert Shapley
bottom panel). Note that this manipulation makes needed in order to distinguish between these two
discriminating the shapes of the high-curvature possibilities (or to show the involvement of both).
stimuli of the ‘‘train’’ block even easier than before. The next question we asked was whether the
Would exposure to these stimuli lead to robust ‘‘train’’ block had to contain high-curvature stimuli,
learning? The results are shown in figure 13.7 (mid- or whether learning could be induced with other,
dle panel on top) in terms of threshold performance ‘‘easy’’ stimuli. This question addresses a possible
as a function of block type. For comparison, we in- interpretation of the abrupt learning observed in sec-
clude the results of ten subjects who participated in tion 12.2, which is that the introduction of the high-
the first session of the experiment described in sec- curvature stimuli allowed the subjects to establish
tion 13.2, where the ‘‘test’’ and ‘‘train’’ stimuli had two distinct categories, or templates, for the ‘‘thin’’
the same support ratio (group A, left panel on top). and ‘‘fat’’ surfaces. According to that interpretation,
It is evident that, whereas all the subjects in group A the minute di¤erences in curvature given in the first
improved in the ‘‘train’’ block and retained their (‘‘test’’) block were not enough to establish two such
learning in the ‘‘retest’’ block, the subjects of group distinct categories, and this led to the poor perfor-
B showed large individual di¤erences in their per- mance observed. Once subjects were able to form
formance. Moreover, the performance of even those the categories, using the exaggerated examples given
subjects who improved during the ‘‘train’’ block fell in the ‘‘train’’ block, they were able to classify the
back to its initial (‘‘test’’) level in the third, ‘‘retest’’ low-curvature stimuli correctly, too. This interpre-
block. We conclude that the improvement in per- tation suggests the following prediction: significant
formance observed in group A, and the accompany- improvement should not be observed when the
ing transition in the perceptual organization of the ‘‘train’’ block is changed so that large-curvature
small-curvature (‘‘test’’) stimuli into illusory surfaces, stimuli are no longer given. This prediction, how-
can only be triggered by large-curvature IC stimuli ever, was not supported by the following ex-
with similar size inducers (i.e., with the same support periment. A third group of subjects (group C) was
ratio). One reason for this result may be that illusory given three consecutive experimental sessions, where
contours of di¤erent support ratios are generated the second (‘‘train’’) session consisted of only low-
or represented by di¤erent neural substrates, even curvature (1–3 ) stimuli. To facilitate performance in
though the illusory surfaces themselves look percep- this block, two sets of long-duration low-curvature
tually similar (e.g., di¤erent neurons respond to the stimuli were added to the stimulus set: 153 msec
local occlusion cues, or L-junctions, as the support (þ 69 msec blank screen) and 194 msec (þ 83 msec
ratio is changed, because those junctions fall on dif- blank screen; the rest of the 1–3 stimuli had expo-
ferent retinal locations). Alternatively, the lack of sure durations of 97 msec þ 69 msec blank screen, as
learning observed in group B may be related to in the ‘‘test’’ block). In other words, what made the
cognitive factors: the mixture of ‘‘very easy’’ and additional stimuli in the ‘‘train’’ block easy this time
‘‘very hard’’ stimuli that are easily discriminable (due was that they had, not higher curvature, but a much
to the di¤erent support ratios) in the same block may longer exposure duration. The results of group C are
have led to a di¤erential treatment of the two sets of presented in figure 13.7 (right panel, top) in terms of
stimuli by the subjects. Further experiments will be threshold performance as a function of block type,
The Role of Insight in Perceptual Learning 247
Figure 13.8
Trial-by-trial analysis of percentage correct (left-hand panels) and reaction time (right-hand panels) of group C, who
received a ‘‘train’’ block that contained long-duration low-curvature stimuli, and no high-curvature stimuli (see figure 13.7,
bottom panel, bottom row). Abrupt improvements, similar to those observed for group A (see figures 13.5 and 13.6), are
seen at the transition (trial 13) from the ‘‘test’’ to the ‘‘train’’ block.
improvement did not generalize to a new retinal The fact that abrupt (or insightful) and stimulus-
size, and retraining was necessary for the good per- specific improvements can happen within the same
formance to reoccur. The onset of the learning also experimental paradigm suggests that there may be a
showed great sensitivity to the spatiotemporal prop- connection between the mechanisms that underlie
erties of the training stimuli, again demonstrating a these two forms of learning, which were previously
strong perceptual component in the learning. thought of as separate. One implication of this view
The paradigm presented here provides a unique is that perceptual learning should be thought of as an
situation in which these two properties of the active process, where the subject’s continual e¤ort to
improvement—abruptness and stimulus specificity— process the incoming sensory information in the
occur together. Usually, they tend to characterize most e‰cient and meaningful way is crucial for the
quite di¤erent forms of learning. Stimulus specificity improvement to take place. According to this view,
has been found mostly in cases where the learning the fact that abrupt improvement can be induced
was gradual and incremental—often requiring hun- experimentally should be viewed as a manifestation
dreds or even thousands of trials (Ramachandran and of the underlying active process of exploration on
Braddick 1973; Ramachandran 1976; Fiorentini and the part of the subject, a process that is taking place
Berardi 1980; Ball and Sekuler 1982; Karni and Sagi continuously. Indeed, studies by Shiu and Pashler
1991; Ahissar and Hochstein 1993, 1997; Masson (1992) and Ahissar and Hochstein (1993; see also
1986; but see also Karni and Sagi 1993; Poggio, chapter 14, this volume) provide further evidence
Fahle, and Edelman 1992). Insight, on the other for this idea. In their experiments, an identical set of
hand, involves a sharp improvement by its very na- visual stimuli could be presented in the context of
ture. Moreover, the name ‘‘insight’’ itself suggests two di¤erent tasks. They found that the extensive
that the subject has found some new ‘‘solution’’ to exposure that gave rise to the improvement in the
the problem at hand—a new understanding of how ‘‘main’’ (trained) task a¤ected performance in the
to perform the task or solve the problem. It implies other (untrained) task very little. One exception to
that we should not expect the improved perfor- this finding, however, may be when the task is
mance to be dependent on factors such as the con- ‘‘preattentive,’’ in the sense of showing little or no
text (e.g., in the case of problem solving) or the performance loss as the attentional load is increased
retinal size or location of the stimulus (in the case of by enlarging the stimulus array (Treisman and
a visual task). Research in problem solving, however, Gelade 1980) or by introducing another, concurrent
indicates the this expectation is not always met: in task (Braun and Sagi 1991). Ahissar and Hochstein
fact, subjects can show great susceptibility to the (1993) found a significant amount of improvement
surface-level attributes of a problem they learned to in a popout task—a classical preattentive task—after
solve, transferring the solution to another problem subjects received extensive exposure to popout
that shares these attributes, but failing to transfer it to arrays in the context of a di¤erent task. Using a tex-
a problem that has an identical deep structure but a ture segregation task, Karni and Sagi (1993) dis-
di¤erent surface-level structure (Gick and Holyoak tinguished two learning phases. The initial (‘‘fast’’)
1980). These findings echo those reported here, that phase, takes place over several hundreds of training
an improvement which seems to involve an ‘‘insight- trials (in their study this initial phase was associated
ful’’ solution can be stimulus specific. with a drop in thresholds of more than a factor of
The Role of Insight in Perceptual Learning 249
two). Karni and Sagi (1993; see also Sagi and Tanne incremental improvements more likely than large,
1994) proposed that this phase involves top-down sudden ones. This observation was made already by
control and involves the establishments of con- Hebb (1949, p. 160), who noted that in order to
nections that make the task automatic. The second, induce insight, one needs ‘‘tasks . . . of just the right
much slower phase of learning (which takes place degree of di‰culty . . . [the task] must neither be so
over days and led to a further drop in thresholds of easy so that the animal solves the problem at once,
30–40% in their study), is therefore hypothesized to thus not allowing [experimenters] to analyze the
be taking place in a passive, bottom-up way, requir- solution; nor so hard that the animal fails to solve it
ing no active e¤ort on the part of the observers. except by rote learning in a long series of trials.’’
The idea that incremental, stimulus-specific Thus, while previous perceptual learning studies were
learning and abrupt, insightful improvements may not intended to optimize conditions for insightlike
be part of a common learning mechanism implies improvements to occur, it may well be the case that
that it should be possible to show a continuous tran- an appropriate change in the experimental procedure
sition between these two forms of learning. There could promote such abrupt learning in other tasks as
is evidence that such a continuum can indeed be well (e.g., by using the method of constant stimuli to
observed. In the course of performing pilot experi- give a set of di‰cult stimuli followed by the same
ments before those reported here, we ran a large stimuli mixed in with easier ones, as we have done
group of subjects ðn ¼ 34Þ on variations of the para- here). This in turn suggests that, by appropriate choice
digm described in this chapter. Our purpose was to of the stimulus parameters and experimental proce-
characterize the distribution of performance across dure, perceptual learning may be used as a model for
our subject population, in order to optimize stimulus studying insight.
conditions for abrupt learning. By changing the ex- The resemblance between insight in problem
posure duration of all stimuli, we varied the overall solving and in perception has recently also been
level of di‰culty of the task while at the same time noted by researchers writing about the psychology of
maintaining the ‘‘test-train-retest’’ structure reported insight. Schooler, Fallshore, and Fiore (1995) found
here (low curvature in the ‘‘test’’ blocks; mixed high- a strong correlation between insight in problem
and low curvature in the ‘‘train’’ blocks). We found solving and the capability to find the shapes of ob-
that, for shorter exposure durations than those jects in blurred pictures. Gruber (1995) specifically
reported here (i.e., when the task was more di‰cult), drew attention to similarities in the process of inte-
subjects often did not show an abrupt improvement at gration of fragmented images such as illusory con-
the transition from the ‘‘test’’ to the ‘‘train’’ block, but tour stimuli and the integrative processes of insight
instead showed a slower, more gradual improvement in problem solving.
(and sometimes did not improve at all within the ses- Another issue raised by the results presented in this
sion). These results also shed light on the issue of why chapter is the interpretation of stimulus specificity.
traces of abrupt or insightful improvements were not The lack of generalization of perceptual learning to
reported previously in perceptual learning studies. To new stimulus parameters has been previously taken
allow for a substantial amount of improvement, to imply that the learning occurred in early, reti-
researchers used parameter regimes that made their notopically organized visual cortical areas, which are
tasks extremely di‰cult, and thus also made gradual, known to encode position, local orientation, and
250 Nava Rubin, Ken Nakayama, and Robert Shapley
similar attributes at the level of individual cells provement we observed, it would require the simul-
(Ramachandran and Braddick 1973; Fiorentini and taneous modification of synaptic e‰cacies between
Berardi 1980; Karni and Sagi 1991; Poggio, Fahle, multiple (neighboring) cells. The existence of neural
and Edelman 1992; Weiss, Edelman, and Fahle mechanisms that could support such a ‘‘cooperative’’
1993; Fahle 1997; Ahissar and Hochstein 1997; see form of synaptic plasticity is not presently known.
also chapters 9–11, this volume). Placing the site of Thus the abruptness of the learning, combined with
plasticity at an early visual cortical site was consistent the global nature of the task make it unlikely that a
with two notable characteristics of the tasks and the model based exclusively on quick synaptic mod-
learning course. First, the tasks were of a local na- ifications of local connectivity in early cortical areas,
ture, involving interactions between image points 1 such as has been suggested previously for other tasks
apart or less (Ramachandran and Braddick 1973; (Poggio, Fahle, and Edelman 1992), could work for
Karni and Sagi 1991; Poggio, Fahle, and Edelman the phenomenon presented here. The fact that the
1992; Fahle 1997; Ahissar and Hochstein 1997). abrupt learning we observed was specific to retinal
Second, the improvement was incremental, often size indicates, however, that the site of plasticity can-
taking place over hundreds or even thousands of not be limited to higher visual areas that encode
trials (Ramachandran and Braddick 1973; Fiorentini shapes in a size-invariant way, either. It is therefore
and Berardi 1980; Karni and Sagi 1991; Fahle 1997; di‰cult to conceive of the learning as occurring at a
Ahissar and Hochstein 1997). Even where fast single site. Instead, the improvement we observed is
learning phases were observed (Poggio, Fahle, and more consistent with changes in processes that in-
Edelman 1992; Karni and Sagi 1993), performance volve interactions between multiple levels of repre-
showed a steep, but gradual improvement, over sentation of the stimuli, where activity in early visual
several dozens of trials. In contrast, in the experi- areas is a¤ected by stimulus-driven processing as well
ments reported here, there were large retinal dis- as top-down control (Edelman 1987; Grossberg
tances between the inducers (more than 10 visual 1987; Ullman 1995; Dayan et al. 1995; cf. chapters
angle at the 60 cm viewing distance), which means 18, 20).
that the relevant information was stored in widely To conclude, we have shown evidence that in-
separated neurons in early visual cortical areas. This sightlike improvements in performance can take
in itself does not preclude models that assume that place in perceptual learning, and that the improve-
the visual processing required to perceive the illusory ment may show stimulus specificity similar to that
contours takes place in those early, small receptive described before for more incremental, gradual
field areas because information can propagate in a learning. Our results suggest that the distinction be-
few iterations across several relays of lateral con- tween insightful and gradual, incremental learning
nections (Gilbert and Wiesel 1983; Gilbert et al. may need to be revised. Rather than postulating two
1996; Lund 1988; Malach et al. 1993; for a model distinct mechanisms for the two forms of learning,
that makes use of such lateral connections to detect our findings may be better understood within a sin-
global shapes, see also Sha’shua and Ullman 1988). gle framework. This view was put forward already
But here is where the abrupt nature of the learning by Hebb (1949), who wrote that ‘‘insight . . . con-
we observed comes into play. For a model based on tinually a¤ects the learning of the adult animal’’ ( p.
lateral connections to exhibit the kind of sharp im- 163), and that ‘‘it is not wholly separate from rote
The Role of Insight in Perceptual Learning 251
Acknowledgments
the decision-making stage, regardless of whether an they do not interest our attention equally.’’ Re-
overt response is made or planned (Pashler 1998). cently, Hikosaka, Miyauchi, and Shimojo (1993,
Proponents of early selection have drawn on find- 1996; but see Downing and Treisman 1997) have
ings that we can easily select on the basis of simple attributed a newly discovered motion illusion to the
features but not on the basis of complex concepts. facilitatory e¤ect of attention. In this illusion, a line,
For example, Treisman and Gelade (1980) and Julesz which is presented physically all at once, is perceived
(1981) showed that detecting (selecting) on the basis to be drawn from one side to the other, when at-
of a simple ( primitive or early-encoded) feature (e.g., tention has been captured by the first side due to a
intensity, color, orientation) does not require atten- preceding visual cue.
tion, whereas searching (selecting) on the basis of a The ‘‘early versus late’’ dispute that occupied at-
feature combination (conjunction) does. Treisman tention research by cognitive psychologists also dom-
and Gelade found that, without su‰cient attention, inated the rare electrophysiological studies seeking
false conjunctions are perceived between basic fea- the neural basis of attention. These studies measured
tures and elements. For example, when subjects are the e¤ect of the behavioral task performed by a mon-
asked to report what they have seen at positions to key on the response characteristics of single neurons.
which attention has not been focused in a brief dis- Initial findings suggested that the responses of neu-
play containing a green X and a red square, they are rons in early visual (cortical) areas are determined by
likely to report X, square, red, and green. But they the retinal input (e.g., Wurtz, Goldberg, and Rob-
will often confuse the conjunction between color inson 1982). Behavioral e¤ects were only found at
and shape, reporting a red X and a green square. The higher areas along both the dorsal ( parietal lobe) and
finding of illusory conjunctions between basic fea- ventral (temporal lobe) processing streams. More re-
tures was the basis of Treisman’s feature integration cent studies (e.g., Motter 1993) indicate that when
theory, which moved the dispute of early versus late task performance requires selective spatial attention
selection toward understanding the mechanisms un- (such as when there are several elements in the dis-
derlying selection. The theory characterized atten- play), responses are modified by attentional demands
tion as the glue that binds basic features present at a in the primary visual area (V1) as well as in higher
single position (e.g., the orientation and the color of areas (V2, V4).
a light bar) to a unified percept. Interestingly, these studies assume that attention
What is the mechanism through which selection is follows the hierarchy revealed by bottom-up ana-
implemented? One answer is facilitation—the at- tomical connections and neuronal response charac-
tended information is processed faster. This idea was teristics. That is, the earliest site of selection is the
already suggested by James, and is largely accepted primary visual area, where basic features are repre-
today (though still in dispute). Based on his personal sented. Thus, if attention operates at the level fol-
observations, James (1890, 409) suggested that at- lowing feature representation, it operates just after
tention facilitates access to consciousness: ‘‘The smith area V1 ( Julesz 1990). A late site of selection would
may see the sparks fly before he sees the hammer be within inferotemporal areas (considering the ven-
smite the iron. There is thus a certain di‰culty in tral stream as an example), where objects may be
perceiving the exact date of two impressions when represented (see Desimone and Ungerleider 1989 for
The Role of Attention in Learning Simple Visual Tasks 255
14.2.2 Can Learning Occur at an Early Cortical Hochstein 1993, 1995, 1996b, 1997). In these
Site? studies, we considered the stimulus characteristics
for which we found learning specificity to indicate
What type of neural plasticity underlies this dramatic the cortical site of learning. This ‘‘psychoanatomy’’
behavioral improvement? During the 1960s, such an mapping assumes that learning specificity reflects the
improvement would necessarily have been attributed response selectivities of the modified neurons. Two
to a high-level cortical area because low-level areas major conclusions emerged from these studies:
were assumed to lose their plasticity following a
1. Learning can be very specific for basic visual
critical period of development. We now know that
dimensions, suggesting that learning can take place
primary sensory areas may be modified in adults
at very low-level cortical sites where neurons are
(e.g., Gilbert and Wiesel 1992). But which neural
selective for these basic visual dimensions.
site and mechanism should serve as initial primary
candidates for the modifications that underlie learn- 2. The degree of specificity to the trained stimuli
ing in the search for orientation features? varies greatly across subjects and task conditions,
Our initial hypothesis followed a simple line of suggesting that the learning site is not fixed.
reasoning: (1) learning must occur at a site where the The stimulation features that were tested for specif-
task is processed (or represented); (2) odd orienta- icity include stimulus retinal position, stimulated
tions are detected by processes at early cortical sites eye, and stimulus size and orientation. In general,
(as indicated by its being performed automatically the scales of the specificities found were such that do
and with parallel processing; see above); and (3) not allow precise localization of a single underlying
learning of orientation detection may therefore cortical area (see also chapters 8–13). After training
occur at an early cortical site. Our working hypoth- with the target presented at a consistent retinal po-
esis was that the site could be the earliest stage that sition, there was little improvement in performance
explicitly codes for the task, that is, the earliest stage with the target at other retinal positions (Ahissar and
where neurons are selective for orientation (see for Hochstein 1995, 1996b; but see also qualified be-
example, Karni and Sagi 1991). This hypothesis low). Substantial degradation in performance was
yielded several predictions, some of which turned found for a 0:7 shift in target (between training and
out not to be true, as discussed in section 14.3. testing sets) position within the array from near to
farther away. Surprisingly, a shift twice as large, from
14.3 Learning Can Occur at Multiple one position right to one position left of fixation (or
Cortical Levels vice versa) did not hamper performance. The transfer
across such distances, which are very large in terms of
14.3.1 Learning Can Be Specific to the Trained the size of receptive fields in area VI, suggests that,
Stimuli under these conditions, learning occurred, not in the
primary visual area, but rather in areas that retain
To decipher the neural substrate underlying percep- cruder retinal position selectivity. Although similar
tual learning, we performed a series of learning results were found for some other tasks (Berardi and
studies using the feature search task (Ahissar and Fiorentini 1987; cf. chapter 9; Karni and Sagi 1991;
258 Merav Ahissar and Shaul Hochstein
Figure 14.2
Orientation specificity to a variety of orientation manipulations in four subjects (L. L., N. B., Y. E., M. C.): rotating (rot)
distractors (cross), rotating (rot) target only (X), and rotating all elements by 30 (square). All maniputations produce a
substantial threshold increment, except left-right mirror reversal (star). Subject Y. E. was trained and tested monocularly
(right eye: downward-pointing triangles; left eye: upward-pointing triangles), whereas other subjects were trained and tested
binocularly. Subjects were presented a fixed 7 7 array, with target in any of central 5 5 position, except corners and
fixation positions for subjects L. L. and M. C.; with target at one of two diagonal positions for subject N. B.; and with target
as described in figure 14.6 for subject Y. E.
see chapter 10; Schoups, Vogels, and Orban 1995; also observed in all other reported studies that
see chapter 5), large degrees of spatial transfer (e.g., looked for it (Poggio, Fahle, and Edelman 1992;
Treisman et al. 1992) and huge variability between Fahle and Edelman 1993, Beard, Levi, and Reich
di¤erent subjects (e.g., Fahle 1994; see chapter 11, 1995; Ramachandran and Braddick 1973; Shiu and
Beard, Levi, and Rich 1995) were also observed Pashler 1992; Schoups, Vogels, and Orban 1995;
using similar experimental procedures. Training with Fiorentini and Berardi 1981; Treisman, Vieira, and
one eye and testing with the other, we did not find Hayes 1992; Karni and Sagi 1991; Polat and Sagi
ocular specificity (figure 14.2; see Ahissar and 1994), although large between-subject variations for
Hochstein 1996), although some studies have (Karni orientation specificity were also found, as illustrated
and Sagi 1991; but see also Schoups and Orban 1996). in the distribution of orientation specificity across
We observed specificity for a variety of orientation subjects plotted in figure 14.4 (see also Fahle and
manipulations, including rotation of the target, the Edelman 1993). Size specificity was also observed
distractor elements, or both, or swapping their ori- (Ahissar and Hochstein 1993, 1996b; see also Fior-
entations (figures 14.2 and 14.3; see Ahissar and entini and Berardi 1981) together with a surprising
Hochstein 1993, 1996b). Orientation specificity was asymmetry: no transfer occurred from big to small
The Role of Attention in Learning Simple Visual Tasks 259
Figure 14.4
Distribution of percent specificity for 90 rotation across
subjects. For each subject, the percent specificity is the in-
crement in threshold induced by orientation manipulation
Figure 14.3
relative to the total di¤erence in threshold between first
Learning and specificity for swapped orientations. The
and asymptotic sessions. Note that the distribution is not
graphs show examples for three subjects (O. L., E. I., E.
normal. Data were collected from fourteen subjects trained
L.), and average for ten subjects (lower right) who trained
with the task version illustrated in figure 14.6.
with original orientations (triangles) and were subsequently
tested and retrained with swapped target and distractor ori-
entations (circles). Following several training sessions, sub- 14.3.2 Learning Does Not Interfere with Seemingly
jects were retested with the originally trained orientations, Conflicting Stimuli
to examine whether there was interference (final triangle).
Note there is nearly no transfer to the swapped orientations
condition (first circle), that is, learning had to begin again. One training condition consistently yielded almost
Furthermore, learning for the swapped condition is some- complete specificity to the trained orientations. In
what slower than the original learning. On the other hand, this condition, the target could be in any of 48 posi-
after training with the swapped orientations, there was no tions of a 7 7 array (all element positions excluding
interference when retesting with the original orientation
fixation). Despite the uncertainty of target position,
condition—final triangle has just as low a threshold as
asymptote before swapped orientation training—a finding training led to a dramatic improvement in perfor-
contrary to the predictions of the hypothesis of mod- mance, which was extremely specific to both target
ifications at the earliest site. and distractor orientations. We asked whether the
mechanisms underlying learning in this case involved
stimuli, but a large transfer was apparent from small neuronal changes within the earliest level that con-
to big (Ahissar and Hochstein 1996b). tained orientation-selective neurons (Ahissar et al.
In summary, among subjects trained with a con- 1998). Such a mechanism would be simple, and
sistent set of stimuli, learning is largely specific for would have the advantage of not needing a very so-
position, orientation, and size (and sometimes for the phisticated selective teaching mechanism (e.g., prior
eye trained), indicating a low-level cortical site for connections could be strengthened with training
learning. However, the large variability of these experience).
specificities suggests that the site of learning may also Our rationale was that several types of changes
be variable. within the earliest level with orientation selective
units could increase the salience of the odd element.
260 Merav Ahissar and Shaul Hochstein
We rejected these changes one by one, as follows: these stimuli would induce conflicting directions of
modifications (for a description of a specific model
1. An increase in target detector sensitivity. If
having this characteristic, see Peres and Hochstein
learning involved only this change, improvement
1994). We found, however, that none of these
would be selective only to the orientation of the
predictions proved true. Initial performance with
target and would not depend on the orientation of
swapped orientations (following original training)
the distractors. We found, however, that learning
was not worse, and even somewhat better than naive
was specific for both orientations.
performance. Optimal asymptotic performance was
2. An increase in lateral inhibition among the dis- reached in both sets of orientations without inter-
tractor elements. The odd element would become ference, although learning of the swapped set was,
more salient when background activity was reduced. for some subjects, slower than learning of the first set
If learning involved only such a change, then im- (as illustrated in figure 14.3).
provement would only be selective to the orienta-
tion of the consistently trained distractors and would These results refute the simplest, earliest-level hy-
not depend on target orientation because target de- pothesis, indicating that the bulk of improvement in
tector activity would be constant and target salience orientation detection does not stem from changes in
would be just a by-product of interdistractor inhibi- lateral connections between simple orientation de-
tion and its augmentation. We found, however, that tectors. Changes underlying learning must occur at
though learning was specific to the orientation of the levels receiving their inputs from these earliest stages.
distractors, it was even more specific to that of the Yet refuting a specific simple hypothesis does not
target. indicate where learning does occur; indeed, it leaves
many alternatives open. A more specific theory for
3. An increase in lateral interactions between target
choosing the site of modifications underlying be-
and distractor elements, by increased facilitation
havioral improvement is discussed below.
from distractor detectors onto target detectors, by
increased inhibition in the reverse direction, or,
14.3.3 Task Di‰culty Determines Learning
alternatively, by a combination of changes 1 and 2
Specificity
above. If learning involved only these changes, the
orientation of the trained target would have to be
While searching for the source of stimulus specificity
more salient than the orientation of the trained dis-
and learning-site variability, we found that task dif-
tractors. Thus, if we changed the array so that the
ficulty plays a crucial determining role (Ahissar and
orientations were swapped—the target assumed the
Hochstein 1997). In a series of studies, we systemat-
previous distractor orientation and vice versa—initial
ically manipulated the di‰culty of the orientation
performance should be even worse than in that of
detection task for di¤erent subject groups. Task dif-
the naive state. Moreover, there should be no way
ficulty was modified by training with di¤erent ori-
that subjects could maximally enjoy training e¤ects
entation gradients (target-distractor di¤erences of
for practice both with one set of orientations and
90 , 30 , and 16 ), di¤erent target position con-
then with the swapped set of orientations. Improve-
ditions (target in 1 of 48 possible array positions; or
ment with one set of orientations should interfere
in 1 of 2 easy horizontal positions or of 2 more dif-
and disrupt performance with the other because
The Role of Attention in Learning Simple Visual Tasks 261
ficult diagonal positions), and di¤erent processing reaching down, for more di‰cult conditions, to
times (SOAs). Consistently, in groups that trained levels which receive direct input from the lowest
with more di‰cult conditions, learning specificity level that contains direct representation of the tested
for position and for orientation was larger than in dimension.
groups training with easier conditions, as illustrated
(for orientation) in figure 14.5, panel A. These re-
14.4 The Role of Attention in Early
sults suggested that, for di¤erent groups trained with
Perceptual Learning
essentially the same task but with di¤erent degrees of
di‰culty, modifications underlying learning occur at
14.4.1 Selective Attention Is Necessary for Learning
di¤erent cortical sites.
Using a within-subject design, stimulus specificities
Would learning of this simple detection task occur
at di¤erent processing durations (SOAs) were also
even without task-directed attention? To address this
compared (our training procedure employed blocks
question, we designed a paradigm to ensure that sub-
of di¤erent SOAs for each subject, as briefly de-
jects attended the stimuli, and became familiar with
scribed in the caption for panel A of figure 14.1; see
them, while performing a di¤erent task. We then
Ahissar and Hochstein 1997 for further details). The
asked whether implicit learning occurred, that is,
dependence on SOA followed the same rule as the
whether practicing the di¤erent task with these
dependence on di‰culty stemming from orientation
stimuli improved their performance in orientation
gradient or target position. Increasingly di‰cult con-
detection. To this aim, we modified the stimulus set
ditions (shorter SOAs) activated increasingly specific
so that in addition to odd-orientation detection, an-
learning processes. Figure 14.5, panel B, demon-
other task of similar di‰culty could be trained and
strates this dependence. The di¤erent degrees of
tested using the same set of stimuli (Ahissar and
stimulus specificities shown within each subject sug-
Hochstein 1993; see also Shiu and Pashler 1992;
gest that di¤erent learning sites, dictated by di¤erent
Treisman, Vieira, and Hayes 1992; Fahle 1997). The
degrees of task di‰culty, may also characterize the
additional task required attention to another aspect
learning process within a single individual.
of the same stimuli. Subjects were asked whether the
As noted above, we attribute specificity to recep-
entire array was vertically or horizontally aligned.
tive field characteristics at the site of task perfor-
Adding this task required modification of the origi-
mance and modification. Because receptive fields are
nal task, such that the array now consisted of either
more selective for basic stimulus dimensions at lower
5 6 or 6 5 elements, as well as being with or
cortical levels, greater specificity is associated with
without an element of odd orientation (the four
lower-level task performance and learning. Thus the
types of stimuli are illustrated in figure 14.6a). All
linkage between task di‰culty and learning specif-
four types of stimuli were equally likely. On the first
icity indicates that easy learning takes place at high
two sessions, each subject was tested (and trained)
cortical levels (which generalize over basic stimulus
with odd-element detection and array orientation,
dimensions), whereas di‰cult learning takes place at
respectively. In the following sessions, half the sub-
low cortical levels. Thus learning may take place at a
jects trained on one task and half trained on the
variety of sites along the visual hierarchy, beginning
other. When they reached a steady performance
at very high levels for easier task conditions, and
262 Merav Ahissar and Shaul Hochstein
Figure 14.5
Linkage between task di‰culty and learning specificity. (A) Comparison of learning specificity under four training para-
digms. Target could appear at any array position (with equal probabilities) except fixation (all: right column) or at one of two
positions (2 pos: left column). Target orientation deviated from that of distractors by 90 , 30 , or 16 (top, middle, bottom
rows, respectively). Hard conditions are 2 pos 16 or all pos 30 and easy conditions are 2 pos 30 or all pos 90 (see inset at
top left). Transfer to new orientations and positions was larger for easy than for hard task conditions. Average threshold is
plotted as a function of thirds of session for each of the four subject groups. Note improvement, and transfer for easy con-
ditions. (Bottom plot begins at the second third of the first session, because subjects barely noticed the target initially;
threshold was 220 msec.) (B) Comparison of relative orientation specificity ( performance decrement after orientation swap-
ping as percentage of total gained improvement) for the four training conditions and di¤erent stimulus onset asynchronies
(SOAs). Specificity for fixed SOA (top) decreases with decreased task di‰culty; specificity for fixed paradigms and within
subjects (bottom) decreases with increased SOA.
The Role of Attention in Learning Simple Visual Tasks 263
Figure 14.6b
Learning curves showing relative threshold change due to
learning the ‘‘local’’ orientation detection task and to sub-
Figure 14.6a sequent training on the ‘‘global’’ array orientation task
Stimulus set of four stimulus types used to test cross-task (group A at left); and vice versa (group B at right). Note
learning transfer. Subjects trained on ‘‘local’’ odd-element that learning in both cases is task specific, though there is
orientation detection learned to di¤erentiate between the some small degree of global to local task transfer.
left and right panels. Subjects trained on ‘‘global’’ array
orientation detection learned to di¤erentiate between
was governed by attentional selection. Thus sub-
upper and lower panels.
stantial top-down e¤ects must influence low-level
representations (Tsotsos 1990; Nakayama 1991;
level, all subjects were tested and trained with the Motter 1993; Ahissar and Ahissar 1994; Rosenthal
complementary task. and Hochstein 1994; Ishay and Sagi 1995; Ullman
We found that practicing orientation detection did 1995; Treisman 1996; see chapter 20).
not a¤ect subjects’ performance on the array align- The near absence of learning when subjects did
ment task. Practicing the array orientation task im- not attend to the stimulus aspects that were relevant
proved subjects’ performance on the detection task, to task performance in simple feature detection tasks
but only to a small extent. In both cases, further constitutes strong evidence against the possibility of
practice with the new task improved performance learning irrelevant cues. Although one example
substantially. These e¤ects are illustrated in figure cannot prove that such learning never occurs, we
14.6b. are not aware of any condition where learning un-
Our results indicate that attention is essential for attended features has been unequivocally demon-
learning even simple tasks. It is important to note strated. On the contrary, data that were previously
that the role of attention in learning these two tasks interpreted as indicating learning without attention
was critical and robust. Whereas stimulus specificity have recently been reinterpreted. For example, pre-
characterized only some subjects under some con- vious findings on learning rules without explicit
ditions, task specificity characterized all trained sub- instruction to attend to these rules were often inter-
jects. The finding that learning odd-orientation preted as indicating unselective (automatic) learning
detection can be specific to both basic stimulus spa- of deep and global structures. However, it was re-
tial parameters and the attended task suggests that cently shown that improvement may only be found
learning occurred at an early cortical stage and yet
264 Merav Ahissar and Shaul Hochstein
when subjects can use what they are learning while ence is not su‰cient, but is it necessary? Suggesting
they are learning it. Thus, before the training phase, the contrary may seem strange, having found that
subjects were told to attend the presented items, and learning may be quite specific to the trained set of
after the training but before the test phase, they were stimuli, that is, that no learning is found for stimuli
told they were now to determine legality. Therefore, that were not presented. Indeed, stimulus presenta-
they could have used deep structure as a memory aid tion is essential, but perhaps its importance does not
during the training phase, or they could have (in- stem from the need for bottom-up activation. Its
tentionally) analyzed the memorized sequences to importance may stem from the top-down search
figure out their deep structure during the test phase. mechanism needing guidance. Sorting out these two
Indeed, if subjects are not instructed to judge legal- alternatives is di‰cult. For example, we found, (as
ity, but implicit recollection is tested, for example, described above), that when subjects practice the
by asking about ‘‘pleasantness’’ of the stimuli, they do detection of a target presented near fixation, they do
not recognize (i.e., consider as pleasant) novel items not improve in detection of targets farther away
that follow the same deep structure as the test set. from fixation. This position specificity may stem
Thus there is no evidence for implicit learning of from the need for bottom-up activation to induce
the rules when these are irrelevant (see Wright and learning at this position. Alternatively, it may result
Whittlesea 1998). Another phenomenon, priming from selective attention being allocated to this posi-
for spelling homophone words presented to the tion, in which case, if we can induce subjects to pay
unattended ear, was, on careful reexamination, also ‘‘the right kind’’ of attention to a position where no
eliminated (Cowan and Wood 1997). These exam- target is presented, their detection at this position
ples imply that while learning may not requires would improve. To test this, we had a group of
awareness, it requires selective attention. Indeed, Bar subjects practice orientation detection with targets at
and Biederman (1999) have demonstrated subliminal one of two positions, two elements to the right or to
priming for briefly presented object drawings. Their the left of the fixation point (Ahissar and Hochstein
subjects were not able to consciously identify the 1996a, 2000). Because attending two separate posi-
objects, yet they actively attempted to identify them tions with no ‘‘attentional link’’ between them was
during both their first and second exposures. Thus, expected to be di‰cult, we reasoned that subjects
performance may improve without awareness of the might attend to a single ‘‘window of attention’’
learning process, but there is no evidence that it can extending from one target position to the other and
occur without selective attention. including the area between them. We found that
Taken together, these data seriously call into ques- learning occurred for the entire area between and
tion whether any purely perceptual task of detection, including the two target positions, but not for other
discrimination, or identification can be learned with- areas of the array (as illustrated by the before and
out aspect-specific attention. after ‘‘learning maps’’ of figure 14.7). Thus learning
occurred where attention was allocated, regardless of
14.4.2 Selective Attention Is Su‰cient for Learning whether a target was ever presented in these posi-
tions. Similar findings for interpolation of improve-
We have seen that attention is necessary for training, ment in a vernier acuity task were reported by Beard
but is it su‰cient? We have seen that stimulus pres- et al. (1996).
The Role of Attention in Learning Simple Visual Tasks 265
Figure 14.7
Detection maps before (left) and after (right) training with target at one of two nonadjacent horizontal positions (denoted by
þ). Improvement a¤ected a single continuous area spanning intermediate positions. Brightness of map region indicates
fraction of detection of targets at that position (see scale at bottom center) in pre- and postlearning sessions. The maps are
measured by pre- and posttraining sessions with the target presented in any position in the 7 7 array.
These findings are consistent with the interpre- 1. Multiple representations. Multiple mechanisms exist
tation that the role of the stimulus in perceptual within the visual hierarchical pathways for perform-
learning is mainly to guide the top-down search ing each perceptual task. Processing within a partic-
mechanism to select the appropriate population. A ular level of the hierarchy is chosen for determining
more specific hypothesis for this selection mecha- the performance of the task. Perceptual learning is
nism is presented below. seen as a gradual refinement of the neuronal popu-
lation selected as the best for implementing the task.
14.4.3 A Unified Hypothesis: The Reverse 2. Reverse hierarchy. Selection of the neuronal pop-
Hierarchy Theory ulation whose activity will determine task perfor-
mance begins at the top of the visual hierarchy and
The well-studied anatomical and physiological hier- proceeds downward. Selective attention chooses
archy of cortical visual areas defines the sequence of activity in higher-level populations (e.g., the infero-
stimulus processing (Van Essen et al. 1990). We pro- temporal cortical areas) as a default level. The ad-
pose that learning begins at high levels and proceeds vantage of these levels is that their representations
downward in reverse direction along this hierarchy are tuned to global entities in the external world,
(see schematic illustration in figure 14.8). The re- such as objects (rather than to local features, as in
verse hierarchy theory rests on three basic concepts lower-level areas). Indeed, we see and attend to
(Ahissar and Hochstein 1997): objects rather than features, unless otherwise specifi-
266 Merav Ahissar and Shaul Hochstein
The reverse hierarchy theory maps behavioral pat- cific. Associating feature search with low-level areas
terns to a physiological framework and consequently and conjunction search with higher areas is at odds
provides a unified terminology. For example, spe- with physiological results indicating small receptive
cificity to basic spatial features was previously inter- fields in lower-level areas and large receptive fields in
preted as stemming from learning at low cortical higher areas. The reverse hierarchy theory claims,
areas. However, some subjects show this specificity, instead, that easy feature search is learned high along
whereas others, trained under the same conditions, the bottom-up hierarchy, but ‘‘early’’ within the
do not. This variability, attributed to strategic dif- top-down hierarchy of attention. Di‰cult conjunc-
ferences, was not addressed by previous psycho- tion search requires more selective spatial attention
physical studies. We now suggest that the broad choosing lower-level populations and consequently
range of specificities may be accounted for within a yielding position-specific learning.
single set of rules, which dictate the pattern of
choosing a site for learning rather than a specific site
14.5 Accounting for a Plethora of Behavioral
chosen for all subjects under given training con-
Findings
ditions. Although the same rules apply to all subjects,
the specific site chosen by the search mechanism
14.5.1 Learning along a Continuum
depends on the saliency of the di¤erence for each
subject’s visual system. Because, given di¤erent ge-
A robust finding in both the laboratory and daily
netics and experience, the saliency of the same
experience is the phenomenon of ‘‘learning along a
stimulus varies across subjects, attaining a specific
continuum’’ described by Pavlov almost a hundred
level of saliency may lead to di¤erent choices of
years ago (see Sutherland and Mackintosh 1971 for
learning site in di¤erent subjects.
review). Pavlov used the conditioned reflex as a tool
Interestingly, the reverse hierarchy theory unifies
for studying the behavior of learning. He found that,
terminology of both stimulus-driven (e.g., phys-
when he induced conditioning with a single pure
iological) processing and attention-driven (e.g.,
tone, the dog salivated also to the sound of neigh-
psychological) processing. Stimulus processing is
boring tones, an e¤ect he called ‘‘stimulus general-
bottom-up and its hierarchy follows a bottom-up
ization.’’ To obtain more specific conditioning, he
sequence, as revealed in anatomical and physiolog-
had to train the dog to di¤erentiate by repeatedly
ical studies. In this respect, area V1 is ‘‘early’’ or
reinforcing (giving a reward for) one tone and not
low level. On the other hand, attention processing
the other. But when these stimuli were very similar,
(which guides learning) follows a top-down se-
the dog could not learn to di¤erentiate between
quence that begins at higher areas. From the atten-
them even after training for a long time. Pavlov dis-
tion perspective, area V1 is ‘‘later’’ or higher-level
covered that it was preferable to begin training with
and is reached only when spatial resolution is
easy discriminations. He therefore first used stimuli
needed. Thus an apparent paradox between inter-
that were very di¤erent from each other and then
pretations is easily resolved. Treisman, Vieira, and
gradually diminished the di¤erence between them.
Hayes (1992) found that learning an easy feature
This type of experimental procedure is called
search transfers across positions, whereas learning a
‘‘transfer along a continuum.’’ One starts with an
di‰cult conjunction search is largely position spe-
268 Merav Ahissar and Shaul Hochstein
easy discrimination, and the easy learning transfers sponse. Thus, if the discrimination required at the
to somewhat more di‰cult discriminations, and so onset of training is too di‰cult, the system will not
on. Since Pavlov’s time, this procedure has been figure out which changes should be made and dif-
repeated often, notably, by Lawrence (1952) and ferentiation will not be achieved. Still, the con-
Sutherland (et al. 1963), who found it has the air of a nections associating between the perceived stimulus
paradox: ‘‘Animals pre-trained with an easy discrim- and the reward will be strengthened. The reverse
ination perform more accurately on a di‰cult prob- hierarchy theory is thus largely an extension of the
lem than animals trained from the outset on the same learning theory proposed by Sutherland and Mack-
di‰cult problem’’ (Sutherland and Macintosh 1971). intosh, whereby ‘‘learning to attend’’ is described in
The surprising aspect is that, to achieve optimal a more concrete framework.
performance with a di‰cult discrimination task, it
would be best not to train with this di‰cult task for 14.5.2 Learning Dynamics and Enabling: The
many trials, even if correct responses are rewarded. It ‘‘Eureka’’ E¤ect
would be best to first give easier, more informative
trials. The reverse hierarchy theory has two specific pre-
Given our daily experience, ‘‘transfer along a con- dictions with respect to learning dynamics:
tinuum’’ is not intuitively surprising. That is, we all
1. Learning will proceed from easy to di‰cult cases
know that easy cases are a good introduction to hard
even when these are presented in a mixed order.
ones. We come to appreciate the paradox when we
Analyzing the dynamics of learning by examining
try to account for this finding in terms of underlying
improvement within each SOA separately revealed a
mechanisms.
cascade of learning (figure 14.9, left): learning at
Learning is traditionally viewed as strengthening
each SOA was sigmoidal. The rising phase started
the connection between representation of the ap-
earliest for easiest SOAs, and only after substantial
plied stimulus and that of the rewarded response.
improvement was obtained for these did improve-
We would thus expect that the greatest stimulus-
ment begin for harder, shorter SOAs. The slower
response strength would be achieved by repeated
improvement for short SOAs reached asymptote
application of the desired stimulus-response asso-
only after more than two sessions (figure 14.9, right).
ciation. Sutherland and Mackintosh (1971) tried to
The finding of a cascade of improvement is strong
account for this seeming paradox by assuming inter-
support for the assumption that learning of di¤erent
mediate stages between those activated by the stim-
degrees of di‰culty occurs in separate neuronal
ulus and those underlying the choice of response.
populations. Consider the alternative, that learning
Di¤erentiation is achieved by learning in stages that
a¤ects di¤erent extents of a single network, depend-
code the stimulus. The system needs to ‘‘figure out’’
ing on training condition. In this case, di‰cult tasks
which connections within these stimulus-coding
(for which we found stimulus specificity) would
stages should be strengthened. On the other hand, the
modify a subpopulation of that changed by easier
repeated stimulus-response association strengthens
tasks (for which there is more learning transfer to
only connections between the output of the stimulus-
new conditions). In the context of our paradigm
coding stages and those representing the specific re-
with interleaved blocks of easy and di‰cult trials, a
The Role of Attention in Learning Simple Visual Tasks 269
Figure 14.9
Dynamics of improvement for each stimulus onset asynchrony (SOA). Average fraction correct is plotted as a function of
trial number for the group practicing with the target anywhere within the array and with target-distractor orientation gra-
dient of 30 . Short blocks (20 trials) of di¤erent SOAs were presented in pseudorandom order so that, on average, various
SOAs were presented within each 100-trial bin plotted here. During the first session (left), performance for each SOA
remained at first near chance (50%), and improvement began only following some practice. The extent of this initial plateau
was SOA dependent, being short (several dozen trials) for long SOAs and long for short SOAs (@400 trials for 50 msec SOA;
> 500 trials for 33 msec SOA). The slope of improvement was also SOA dependent, being steeper for longer SOAs. Toward
the end of the first session, performance stabilized. Further improvement was gained on subsequent sessions, as seen by
comparison to asymptotic performance in the last session (right).
natural outcome of the latter interpretation is that hard cases were trained, most subjects showed no
training easy conditions would induce immediate improvement throughout the session (figure 14.10,
improvement for harder conditions. The temporal left), even though they were shown a schematic
o¤set between improvement for di¤erent degrees of illustration of the stimulus.
di‰culty suggests that modifications occur at sepa- The more interesting prediction of the theory is
rate neuronal levels. that the system need only perceive one easy case to
2. If training includes only di‰cult cases, even sub- highlight the appropriate low-level region. Searching
stantial practice may not yield improvement because for a positive result, we tested another group of sub-
the system will not have acquired a pointer to the jects with the di‰cult case following a long (30 sec)
appropriate lower-level site containing a detailed presentation of the whole array on the monitor
spatial representation. This phenomenon, as de- screen, once with and once without an odd element.
scribed above, had already been noted by Pavlov. We The results for this group were categorically di¤erent:
tested this in the context of orientation detection. A most subjects showed dramatic improvement (figure
group of subjects was trained for a whole session 14.10, right). We called this large, single-presentation
with only 50 msec SOA (without feedback). This enabling of learning the ‘‘Eureka’’ e¤ect (Ahissar and
SOA was chosen because most subjects managed Hochstein 1997). Consistent results were also found
eventually to improve on 50 msec SOA trials sub- in other paradigms involving subjective contours
stantially when they were intermixed with easier (Rubin, Nakayama, and Shapley 1997; see chapter 13)
blocks of trials (figure 14.9). As predicted, when only and texture discrimination (Papathomas et al. 1999).
270 Merav Ahissar and Shaul Hochstein
Figure 14.10
‘‘Eureka’’ e¤ect. Each curve plots performance of one subject. Subjects received verbal instructions and were shown pen and
pencil drawings of the array with and without an odd element. (Left) When presented only with threshold-level trials of
50 msec stimulus onset asynchrony, most subjects failed to improve at all during their first session. Indeed, 16 out of 23 failed
to reach 60% correct even in last quarter of first session (compare figure 14.9, left, triangles, where same 50 msec trials were
interleaved with easier trials). (Right) Following an extended view of the stimulus, once with and once without the odd ele-
ment, most subjects improved dramatically during the first session: 12 out of 16 exceeded 70% correct (dashed lines indicate
subjects who were tested with swapped orientations relative to enabling stimulus, as in figure 14.1, panel A).
14.5.3 The Drawbacks of Spoon-Feeding extensively trained continuum has marked a path
along the search tree leading to the populations best
Intuitively, we think that spoon-feeding is not an suited for the previously trained examples. Subse-
e¤ective training procedure if one wants to encour- quent search may get trapped into following the
age creative, independent thinking. Perceptual train- marked path when seeking backward for appropriate
ing may follow the same rule. Spoon-feeding, in the nodes. Thus finding the best-suited population for
context of the reverse hierarchy theory, is intensive resolving the novel examples may be harder and more
training with explicit examples spanning a whole laborious than initial learning. As in daily examples of
continuum of the easy to di‰cult range. Training problem solving, subjects will more likely choose the
may be e¤ective because subjects can gradually learn previously experienced solution than a potentially
to discriminate between similar examples (see above). more appropriate novel one.
Yet the cost may appear in subsequent training with a This seems to be the case in training orientation
di¤erent, but similar set of examples. Initially, per- detection. Learning along the continuum is e¤ective
formance may gain from previous training. Higher- when training a whole range of SOAs. Later train-
level nodes, used for the previous, similar set, may ing, with swapped orientations, begins with some-
help to resolve the easy cases. But solving the more what better initial performance, but attains a similar
di‰cult cases, which require a more specifically tuned asymptotic performance level and takes longer than
representation, may now be harder to achieve. The training with the original set, as illustrated in figure
The Role of Attention in Learning Simple Visual Tasks 271
14.3. Interestingly, using a single Eureka presenta- improvement was apparent in performing the ori-
tion, thus bypassing some nodes (though not too entation detection task. Only later did improvement
many because that would result in a frustrating con- begin on the detection task (Ahissar, Laiwand, and
dition, with no learning at all), subsequent learning Hochstein 2001). Task learning seems to require
with di¤erent orientations is not slower than the undivided attention. The order of learning for these
original learning set (Ahissar and Hochstein 1997). tasks did not depend on the order of the response.
The reverse hierarchy thus defines guidelines for Half of the subjects answered first ‘‘1’’/‘‘0’’ for odd-
an optimal training procedure. Training should aim orientation presence/absence, and then ‘‘1’’/‘‘0’’ for
for conditions in which su‰cient stimulation is pro- T =L; and the other half answered in the reverse
vided to initiate and direct an appropriate attentional order. Nevertheless, the central letter task took pri-
search. Yet it is better to avoid the full continuum so ority in securing the undivided attention required for
as not to overmark paths that need to be avoided in learning. Odd-orientation detection began to im-
the future when conditions change. prove only after the central identification task was
mastered.
<
pered performance of both tasks (Ahissar, Laiwand,
and Hochstein 2001). This pattern of interference
suggests that, even when dual-task performance is al-
most as good as single-task performance, processing
still requires (limited) attentional resources. In our
case, the task that received priority, either by being
performed first or by receiving more attentional re-
sources, was central letter identification. The source
of this preference may be that the letter was at fixa-
tion, whereas the orientation detection task required
spreading out attention peripherally to a large por-
tion of the array (Hochstein et al. 2000).
In summary, we have tried various methods of
training a simple visual task. Although subjects’ task
performance substantially improved and they needed
dramatically less processing time, their improved
performance was not accompanied by acquired auto-
maticity. Their performance became more e‰cient,
but processing was still governed by the initial atten-
tional limitations.
Acknowledgments
Figure 15.2
(a) Perceptual learning curve showing practice-dependent
improvements on the vernier hyperacuity task. (From
Fahle, Edelman, and Poggio 1995.) (b) Perceptual learning
curve showing nontransference of learning to a di¤erently
oriented vernier task.
Figure 15.4
In conventional conceptualizations of visual processing, information is believed to flow primarily in the forward direction,
from the image to the early visual areas and thence to the higher stations. That it may flow in the opposite direction rep-
resents a radical departure from the conventional scheme (cf. chapter 20, this volume). It is important to remember that the
two schemes are not mutually exclusive. Both patterns of information flow may coexist.
tion being accomplished by early stages, and the cisms that have been leveled against them. Our overall
results being sent on to the later stations concerned aim in this section is to suggest that, despite a rich
with visual cognition (figure 15.4). The possibility history of research, the issue of high-level influences
that learning at the later stages influences the com- on early perception is still contentious and largely
putation of basic attributes can profoundly change open.
this general hierarchical conceptualization. It may
suggest that in addition to flowing from early per- 15.3.1 High-Level Influences on Early Perception:
ceptual stages to the later ones, visual information The Case ‘‘For’’
may also move in the reverse direction.
Prior Expectations
Several studies have shown a strong influence of
15.3 Previous Work prior expectations on the perceptual interpretation
of ambiguous figures. Many of the stimuli used in
Here we review three important past attempts to these studies, such as the wife and mother-in-law
determine whether early perception is shaped by picture in Boring 1930 (see figure 15.5, panel a) and
high-level influences and then summarize the criti- the Dalmatian dog photograph reprinted in Gold-
High-Level Learning of Early Visual Tasks 277
Figure 15.5
Pictures that illustrate high-level influences on image interpretation. (a) ‘‘Wife’’ (left) becomes ambiguous figure (middle),
which then becomes ‘‘mother-in-law’’ (right). (From Boring 1930.) (b) Dalmatian dog picture. (From Goldstein 1996.)
278 Pawan Sinha and Tomaso Poggio
even relatively high-level reinforcement contingen- accounted for completely by simple bottom-up
cies could modify subjects’ perception. mechanisms that incorporate some general-purpose
natural constraints. The usefulness and versatility of
15.3.2 High-Level Influences on Early Perception: bottom-up mechanisms has been demonstrated in
The Case ‘‘Against’’ a few di¤erent domains, such as color perception
(Land’s Retinex model; Land and McCann 1971;
Although the studies reviewed in section 15.3.1 ap- Land 1983) and the perception of structure from
pear to provide strong evidence in support of the motion (Ullman’s use of the rigidity constraint;
idea that early perception may indeed be influenced Ullman 1979). It is thus plausible that perception is
by high-level cognitive processes, several researchers simply a matter of identifying the right natural con-
have challenged them on a variety of grounds. Three straints and embedding them in perceptual mecha-
key criticisms are listed below. nisms that can function in a bottom-up fashion
without requiring high-level, object-specific knowl-
1. The experimental data show changes in feature
edge. One of the most articulate proponents of
grouping rather than changes in the perception of
this idea is Zenon Pylyshyn (1984, 1999), who has
attributes per se. Thus, for example, the change in an
attempted to explain perceptual phenomena believed
observer’s percept from a meaningless collection of
to be due to high-level influences more simply via
blobs to a meaningful scene in the well-known
appropriate bottom-up processes. He has champ-
Dalmatian dog picture may involve, not a funda-
ioned the view that early perception is cognitively
mental change in the basic perceived attributes such
impenetrable, according to which it would be di‰-
as edges and gray levels, but simply a cognitive
cult, if not impossible, to demonstrate any high-level
regrouping of these attributes. For another example,
influences on early perception.
the act of seeing faces in clouds may be based not so
much on a change in the perception of basic visual Thus it is fair to say that, despite a rich history of
attributes as on a cognitive restructuring of the avail- research, whether there exist high-level influences
able perceptual tokens. on early perception is still largely an open question.
2. It is not clear from the data whether responses are In the sections to follow, we describe some recent
driven by a change in cognitive decision criteria or experimental and computational work that, by re-
perception. Experiments most susceptible to this visiting this question, provides stronger evidence
criticism purportedly demonstrate an influence of that high-level learning does indeed influence early
personal attitudes, values, and high-level reward- perception.
and-punishment contingencies on early perception,
yet their results can as easily be accounted for by a 15.3.3 Does High-Level Learning Influence Early
deliberate change in response criteria as by any fun- Perception? Recent Work
damental changes in perception per se.
A few recent studies have provided evidence sug-
3. Simple natural constraints can account for per-
gesting that high-level learning plays a role in dif-
cepts. This criticism is motivated by whether it is
ferent perceptual tasks. Notable among these are
at all necessary to invoke high-level influences to
Peterson and Gibson 1993, 1994, which found that
explain perception, or whether perception can be
280 Pawan Sinha and Tomaso Poggio
Figure 15.9
Two sample three-dimensional shape recovery results ob-
tained by Sinha and Adelson’s algorithm (1993).
Figure 15.11
Two-dimensional images conveying a vivid sense of three-
dimensionality. Computational schemes that employ a
small set of prespecified shape biases have been able to
mimic human 3-D shape perception with images of rela-
Figure 15.10 tively regular geometric objects, such as the two shown on
Sample test images for which Sinha and Adelson’s algo- the left. They prove ine¤ective, however, in more general
rithm (1993) produces perceptually correct three-dimen- domains, which might contain images such as the Matisse
sional shapes. line drawing shown on the right. (From Sinha and Poggio
1996.)
structures. During the training phase, to allow them with ten di¤erent objects, each object being used
to observe the correlation between the object’s mean only once to prevent any given test session from
angle projection and its associated 3-D structure via serving as the training session for a later trial. To
the well-studied kinetic depth e¤ect (KDE; Wallach determine whether nonrigidity would be perceived
and O’Connell 1953; Ullman 1979), subjects were even in the absence of any previous perceptual ex-
shown one such object rocking through an angle of perience with a given object, members of the second
G20 about a frontoparallel horizontal axis passing (control) population did not undergo training ses-
through its centroid (the ‘‘training’’ object). The test sions for any of the ten objects used. The procedure
phase commenced five seconds after the training ses- for the third population was identical in all respects
sion and was intended to assess subjects’ shape learn- to the procedure for the first except that the roles of
ing. Subjects were shown either the training object the training and test objects were reversed. The ex-
rocking back and forth or another object having the perimenters expected to see a reversal of the pattern
same mean angle projection but a completely di¤er- of results from population 1 to population 3 if the
ent 3-D structure (the ‘‘test’’ object). They were asked two-dimensional–three-dimensional association was
to indicate whether the objects shown looked rigid truly a function of the visual experience during the
or nonrigid. Sinha and Poggio expected that if the training session.
subjects had indeed learned an association between Figure 15.13 shows the results from the three
the training object’s mean angle projection and its 3- populations. The results are plotted as the percentage
D structure, then, when presented with a test object of trials (over all objects) during which subjects per-
having the same mean angle projection, they would ceived nonrigidity in the object depicted in the
perceptually impose on it the learned 3-D structure motion sequence. Members of population 1 exhibit
and the observed motion pattern would be mapped a clear e¤ect of visual experience ( panel a). While
onto this learned structure. If the test object actually the training object appeared nonrigid for an average
had a di¤erent 3-D structure, the sequence would (across all four subjects) of 7.5% of the presentations,
appear to depict a nonrigid deformation. The ex- the test object (rotating about the same axis) ap-
perimenters expected that switching the roles of the peared nonrigid in about 40% of the presentations.
training and test objects across di¤erent populations Members of population 2, on the other hand, per-
of subjects would lead to reversals in the patterns of ceived both the training and the test objects as being
results observed, and that subjects who had not un- rigid in most of the presentations ( panel b). Mem-
dergone the training session would not perceive bers belonging to population 3 exhibited a reverse
nonrigidity because of the well-known bias toward pattern of results ( panel c) relative to population 1.
rigidity (Wallach and O’Connell 1953; Ullman The marked di¤erences in the results of popu-
1979), possibly arising from visual experience over lations 1, 2, and 3 attest to the strong influence of
evolutionary timescales. visual experience on subsequent perception of three-
The subjects were placed in three nonoverlapping dimensional form. One can conclude that it is possi-
populations. Members of the first population indi- ble for the human visual system to learn associations
vidually underwent a training phase and then a test between arbitrary two-dimensional projections and
session, as described above. Each subject was tested 3-D structures, and that this learning subsequently
High-Level Learning of Early Visual Tasks 285
Figure 15.13
Results from three populations of subjects tested by Sinha and Poggio. Subjects were asked to indicate whether the objects
shown in the motion sequences looked rigid or nonrigid. (a) Results from population 1 on training and test objects. (b)
Results from population 2, who had not undergone any training sessions. (c) Results from population 3, for whom the
training and test objects were switched relative to those for population 1. (From Sinha and Poggio 1996.)
influences 3-D form perception. At least for the each of the five objects) exhibited significant e¤ects
kinds of objects considered in Sinha and Poggio’s of learning upon being tested a day later. Their
study, mutual temporal correlation seems more im- learning also seemed to be object specific in that the
portant for the formation of such associations than percept of nonrigidity declined with the addition of
any intrinsic structural characteristics of the 2-D and two-dimensional positional noise to the vertices of
3-D shapes. Supporting this interpretation, subjects the original projection ( panel b), although trans-
reported recognizing the test object as being the one formations that preserved the shape of the training
they had seen during the training session. projection, such as image scaling, did not have a
To better characterize subjects’ shape learning, the large e¤ect on subjects’ responses ( panel c).
experimenters tested their performance under a few Figure 15.15 shows two illusions devised by Sinha
additional conditions. As figure 15.14, panel a, and Poggio related to the ideas presented above.
shows, the learning is long-lasting. Two subjects Panel a shows a situation in which a completely rigid
from population 1 who were trained for ten minutes wire-frame object resembling a person, when rocked
on each of five objects (five training sessions of two- around the vertical axis (sequence shown in panel b),
minute sequences evenly spaced over four hours on is perceived as a person walking (a nonrigid inter-
286 Pawan Sinha and Tomaso Poggio
Figure 15.15
Two illusions devised by Sinha and Poggio to illustrate object-specific influences on the perception of kinetic depth e¤ect
(KDE) sequences. (a) This rigid structure, when rocked back and forth through an angle of 90 about the vertical axis, is
perceived as a human walking (a nonrigid interpretation). (b) Frames from the motion sequence of structure in panel a.
Mismatch between reality and perception is probably due to the recognition of the figure as a human, which leads the visual
system to interpret the motion patterns in terms of how the human structure is expected to change, an expectation likely to
have been learned through visual experience. The same vertex set, when not recognized as a human (say, when the vertices
are connected up randomly) appears rigid upon rotation. (c) Object that projects to a cube (from a specific viewpoint) but has
a very di¤erent three-dimensional structure. (d ) Motion sequence of object in panel C, with cubelike projection, appears to
depict a highly nonrigid object. When the vertices are connected randomly, the percept of rigidity is restored. Just as for the
structure in panel a, the perceived nonrigidity here is likely to be due to the mismatch between the observer’s expectations
and the presented sequence. It is not clear whether the shape expectation for a cube (and other simple geometric objects) is
learned during an observer’s lifetime or is innate. (From Sinha and Poggio 1996.)
288 Pawan Sinha and Tomaso Poggio
specific learning. Thus tasks conventionally assumed pletely random) to 5 (completely human). Inter-
to be low level and hardwired might, at least in part, spersed with these sequences, truly random and truly
be learned through visual experience. human (unscrambled) sequences were also presented
to the subjects. As shown in figure 15.17, at a
Experiment 2 viewing position parallel to the depth axis depth-
Additional evidence showing high-level influences scrambled motion sequences were rated as high
on early three-dimensional shape perception pro- (4.54; standard error: 0.06) as unscrambled sequences
cesses comes from Bültho¤, Sinha, and Bültho¤ (4.59; standard error: 0.07) by twenty-two subjects.
1996 and Bültho¤, Bültho¤, and Sinha 1998 (see The random sequences were rated much lower
chapter 16). These studies di¤er from Sinha and (2.45; standard error: 0.10). What caught the ex-
Poggio’s in two respects. First, they used dynamic perimenters’ attention was the seeming indi¤erence
nonrigid 3-D objects rather than rigid ones and sec- on the part of the observers about whether they were
ond, they examined object-specific influences on viewing a depth-scrambled or a normal sequence:
stereoscopic depth perception rather than on KDE both such sequences were rated almost identically.
perception. Their basic finding was that subjects’ There could be at least two explanations for this.
high-level expectations about an object’s 3-D struc- Either the subjects were perceptually aware of the
ture could suppress the bottom-up depth informa- depth scrambling, but decided to base their ratings on
tion provided by binocular stereo. 2-D ‘‘goodness,’’ or they were perceptually unaware
The stimuli used by these three experimenters of the depth scrambling. This interesting open ques-
were variants of the biological motion displays tion led the three to design a second experiment to
popularized by Johansson (1973). Specifically, they test for the existence of any recognition-dependent,
were interested in the question of whether three- top-down influences that might serve to suppress
dimensional dynamic objects are represented using information about anomalous depth structure be-
3-D structural descriptions (Biederman 1987; Bie- ing provided by low-level binocular stereoscopic
derman and Gerhardstein 1993) or as two-dimen- processes.
sional motion traces. They started with the premise For use as stimuli, the Bültho¤s and Sinha created
that if the internal representation for biological depth-scrambled biological motion sequences by
motion sequences is largely 2-D (Bültho¤ and Edel- adding six levels of arbitrary depth o¤sets to the tra-
man 1992), then scrambling the depth structure of jectories of the body points, from 0% (no noise
such moving light sequences while leaving the 2-D added) to 200% (a noise level of 100%). These o¤sets
traces unchanged should not adversely a¤ect rec- corresponded to a sequence of randomized depth
ognition performance. The three experimenters positions of the individual points within the depth
constructed a depth-scrambled human walker by bounds of the original walker. The stimulus set also
adding uniform random noise to the depth positions included x-y randomized versions of the depth-
of the points, while leaving their projections scrambled point walkers, which did not convey any
along the depth axis unchanged (figure 15.16). Sub- impression of a human in motion. In each trial, after
jects were shown stereoscopic motion sequences one walk cycle (about 1.5 sec), three points were
of this structure and asked to rate its structural highlighted by red outlines for 1 sec (about two-
goodness as a human figure on a scale from 1 (com- thirds of a complete walk cycle). Subjects had to
High-Level Learning of Early Visual Tasks 289
Figure 15.16
Depth-scrambling a three-dimensional motion sequence of a point walker (top) involves adding depth noise to the joint
positions, while leaving their two-dimensional projections (in the xy-plane) largely unchanged. Stereograms of a single
frame from a depth-distorted point walker sequence (bottom): left pair for cross-fusers and right pair for parallel fusers. (From
Bültho¤, Bültho¤, and Sinha 1998.)
determine whether the three points lay in the same information provided by the early processes of bin-
depth plane. ocular stereoscopic depth perception. They hypoth-
Figure 15.18 shows plots of false alarms (a re- esized that recognition based on two-dimensional
sponse of ‘‘in same plane’’ when the points are in fact traces is accompanied by a strong top-down sup-
not in the same plane) for three conditions of stim- pressive influence that renders observers less sensitive
ulation (see figure caption) for the upright presenta- to anomalies in the depth structure of the presented
tion as a function of depth noise level. The false stimulus. In keeping with its object-specific nature,
alarm rate is highest for human sequences with this influence can be modulated by factors that
points on the same limb. The three experimenters change the recognizability of a stimulus. This hypo-
interpreted these results as pointing to the existence thesis would explain why the false alarm rates for the
of a top-down influence capable of modulating the human sequences are much higher than those for the
290 Pawan Sinha and Tomaso Poggio
Figure 15.18
Figure 15.17 Depth discrimination for upright presentation. Dashed and
Results of goodness rating experiment averaged across dotted curve ¼ human walker with the three marked
twenty-two subjects. Subjects rated the presented se- points on the same limb; solid curve ¼ human walker with
quences of a point walker on a scale from 1 (very random) the three marked points on di¤erent limbs; finely dashed
to 5 (very human). At 0 , the walker is seen walking to the curve ¼ random sequence with three highlighted points.
right with its depth axis parallel to the viewing axis; at 90 , The false alarm rate is plotted against the maximum ran-
the walker is seen walking toward the observer with its dom depth distortion allowed in the sequence. (From
depth axis perpendicular to the viewing axis. Finely dashed Bültho¤, Bültho¤, and Sinha 1998.)
curve ¼ undistorted walker; solid curve ¼ depth-distorted
walker; coarsely dashed curve ¼ walker distorted in x and
nitive decision criteria. It is also important to note
z. (From Bültho¤, Bültho¤, and Sinha 1998.)
that the perceptual processes undergoing change in
these experiments (kinetic depth e¤ect perception in
random ones. Also consistent with this hypothesis experiment 1 and binocular stereoscopic depth per-
was the experimenters’ finding that inverting the ception in experiment 2) have traditionally been
point walkers (an operation that reduces their rec- assigned to early vision, and therefore considered
ognizability while preserving all low-level cues) immune to top-down influences. Having reviewed
reduces the false alarm rate. evidence for the existence of learned high-level
influences on early perception, we turn now to
Summary a recent attempt to computationally model these
Experiments such as the two reviewed above have phenomena.
yielded data that strongly suggest the existence
of learned object-specific influences on three- 15.4.3 A Computational Model for Incorporating
dimensional shape perception. As distinguished from High-Level Learning in Early Perception
studies reviewed in section 15.2, they render the
high-level learning perceptually manifest; in other Let us first consider, in conceptual terms, what such
words, high-level learning changes observers’ per- a model must do. The model’s overall task, to esti-
cepts per se, rather than merely changing their cog- mate perceptual attributes of an object in an image
High-Level Learning of Early Visual Tasks 291
one adopted by Jones et al., is to exploit specific the line drawings were supported on a subset of the
knowledge about faces in order to compute the line pixels of the corresponding gray-level images, the
drawing. This approach runs contrary to the tradi- line drawings associated with novel images could
tional wisdom in computer vision, because it assumes be straightforwardly obtained by warping the line
that object recognition is used for edge detection— drawing associated with the reference prototype us-
almost a complete subversion of the usual paradigm. ing the estimated shape vector (see figure 15.20).
A possible implementation of this approach is based Figure 15.23 shows a few examples of novel images
on a learning metaphor. Consider a set of proto- (not contained in the set of Jones et al. prototypical
typical (gray-level) face images and the correspond- examples) and the line drawing estimated from each
ing line drawings, drawn by an artist. The task is to of them by Jones et al.’s ‘‘ideal edge detector.’’ To
learn from these examples the mapping that asso-
ciates to a gray-level image of a face its ‘‘ideal’’ line
drawing. Computationally, this task is analogous to
the problem of view prediction described above.
Jones et al. (1997) implemented an even simpler
version of the scheme. They assumed that the ideal
line drawing corresponding to the average prototype
is available from an artist, as shown in figure 15.22.
Matching the flexible model obtained from the pro-
totypes (some of which are shown in figure 15.19) to
a novel gray-level image provided a shape vector
that was a linear combination of the prototypes and
that e¤ectively prescribed how to warp the average
shape of the gray-level prototype in order to match
the shape of the novel gray-level image. Because
Figure 15.23
Examples of ideal edges found by the Jones et al. algo-
rithm. The left column shows the input novel images. The
middle column shows the line drawings estimated auto-
matically by the algorithm, which matches the flexible
models to the novel images and then appropriately modifies
the ideal edges of the reference image. For comparison, the
right column shows the edges found by a bottom-up edge
detector (Canny 1986). Note that the ideal edges empha-
Figure 15.22
size the perceptually significant features of the face much
Reference face (left) and corresponding line drawing (right)
better than the Canny edges. (From Jones et al. 1997.)
created by an artist. (From Jones et al. 1997.)
High-Level Learning of Early Visual Tasks 295
contrast this approach to a low-level gradient-based 1991; Mumford 1992; Ullman 1995). We conjecture
approach, figure 15.23 also shows the edges found that visual perception in humans may rely on similar
for each face image by a Canny (1986) edge detec- processes to a greater extent than commonly as-
tor. Figure 15.24 shows the ideal edge estimated for sumed. Of course, biological vision may use bottom-
a partially occluded input image. As is evident from up verification routines to validate the top-down
the examples, Jones et al.’s algorithm can detect and ‘‘hallucination.’’ A similar verification approach (top-
complete edges that do not correspond to any in- down and bottom-up) could also be e¤ectively used
tensity gradients in the image. The power of the in machine vision implementations such as the one
algorithm derives from the high-level knowledge described here.
about faces learned from the set of prototypical Logically, our conjecture consists of two some-
images. what independent parts. First, at least in some cases,
Thus the scheme proposed by Jones et al. (1997) is our visual system may solve low-level vision prob-
an example of a class of algorithms that can be used lems by exploiting prior information specific to the
to learn visual tasks in a top-down way, specific to task and to the type of visual input. And second, vi-
object classes. From the point of view of a neuro- sual systems may learn algorithms specific to a class
scientist, these demonstrations are nothing more of objects by associating in each ‘‘prototypical’’ ex-
than plausibility proofs that a simple learning process ample an ‘‘ideal’’ output to the input view—the
can successfully incorporate object-specific knowl- main thesis of this chapter. The ‘‘ideal’’ outputs may
edge and thereby learn to perform seemingly ‘‘low- be available through other sensory modalities, se-
level’’ visual tasks in a top-down manner (Cavanagh quences of images in time, or even explicit instruc-
Figure 15.24
Examples of ideal edges found for degraded inputs. (a, e) Of the two input face images, the first (a) is partially occluded,
whereas the second (e) has spurious edges such as those caused by patterned shadows. (b, f ) Edges found by the Jones et al.
algorithm for inputs in panels a and e. (c, g) Ideal edges superimposed on the gray-level images. (d, h) Edges estimated by a
bottom-up operator. (From Jones et al. 1997.)
296 Pawan Sinha and Tomaso Poggio
tion. The notion of what constitutes an ‘‘ideal’’ out- decisions rather than modifications of perception per
put corresponding to a certain class of inputs may se (Gibson 1969). Several authors (e.g., Fodor and
change and evolve over time as the learning process Pylyshyn 1981; Pylyshyn 1999) have argued for the
encounters new examples. The second part of our cognitive impenetrability of early perception. Early
conjecture predicts that human subjects should be perception, in their view, is governed by general
able to learn to associate arbitrary outputs to input natural constraints and is immune to learned object-
images and to generalize from these learned associa- specific knowledge.
tions, a prediction confirmed by recent psychophys- We examined this question in the context of
ical evidence, as we saw in section 15.2. The strong a specific visual perceptual task—the recovery
form of this second part is that the learning follows of three-dimensional structures from single two-
the linear combination algorithm Jones et al. used dimensional line drawings, long considered one of
in their plausibility demonstration; the weak form, the key early vision tasks and widely believed to be
which we favor, leaves open the specific learning performed by relying on the use of general natural
scheme. Further work may enable us to verify constraints. We started by describing a constraint-
whether the strong or weak form is to be preferred based shape recovery system and its limitations as a
and if the weak, to determine which learning comprehensive model of the corresponding percep-
scheme is used by the visual system. tual process. Tests of the implemented system sug-
gest that, though it mimics human perception in
specific domains, it does not possess the ability to
15.5 Summary and Conclusions
generalize to other, more naturalistic, domains. All
constraint-based shape recovery schemes proposed
We began this chapter by asking whether learning
thus far, without exception, share this limitation.
shapes perception. Experimental results over the past Although such a limitation certainly does not prove
four decades have provided strong evidence demon-
that the constraint-based account is untenable, it
strating the experience-dependent malleability of
does point to the need for exploring an alternative
perceptual processes (Hubel and Wiesel 1970; Kass
high-level learning-based account for the task.
1994). Several lines of reasoning suggest that the
We next described recently reported experiments
locus of learning in the reported instances is situated
to determine whether the three-dimensional percept
very early in the processing pathway (Karni and Sagi
corresponding to a given two-dimensional image
1991; Ahissar and Hochstein 1998).
can be accounted for in terms of the observers’ prior
Answering the question of whether high-level,
visual experiences. The results suggest that this may
object-specific learning can influence early percep- indeed be the case. Observers can learn to associate a
tion is important for understanding the overall or-
projectionally consistent but otherwise arbitrary 3-D
ganization of information flow in the brain. Early structure with a randomly generated 2-D line draw-
experimental data supporting the position that it ing, with two significant perceptual consequences:
does (Schafer and Murphy 1943; Bruner and Post- misperception of KDE sequences (rigid rotating
man 1947; Bruner and Goodman 1947; Hochberg objects appear highly nonrigid) and misperception of
and Brooks 1958) have been criticized on the stereoscopic depth. The learning is long-lasting and
grounds that such data reflect changes in cognitive specific to particular line drawings. These results
High-Level Learning of Early Visual Tasks 297
provide compelling evidence in support of a role of nation are finally used to compute the desired attri-
high-level object specific learning on 3-D shape butes such as shape corresponding to the input. It is
perception. interesting to note that the information flow being
As a follow-up to these experimental results, we proposed here (image ! recognition ! perceptual
discussed a simple computational model for how attributes) is quite the reverse of what has conven-
object-specific learning may be brought to bear on tionally been assumed (image ! perceptual attrib-
the performance of supposedly early perceptual tasks. utes ! recognition).
Given the identity of an object in an image, or per-
haps just its class membership, the model attempts to
facilitate the computation of basic attributes such as 15.6 Open Questions
three-dimensional shape and contour structure, and
to bring previously learned knowledge about the Recent experimental results have made signifi-
object class to bear on this task. The model relies on cant headway in resolving the question of whether
a prior stage of recognizing the object in an image high-level learning can influence early perception.
but is agnostic about precisely how this may be Interesting computational models have also been
done. It then represents the image as a combination proposed to account for these experimental findings.
of training instances. The parameters of the combi- These e¤orts are bringing us closer to a unification
of cognition and perception. Such a unification—
Figure 15.25
Preliminary object recognition results with ‘‘minimalist’’ representations, quasi invariants constructed from pairwise ordinal
brightness relationships over di¤erent regions of the images. In the examples shown here, the task is to detect human faces
irrespective of the illumination conditions, identity, and complexion. Detections are marked by white spots in the center of
the head.
298 Pawan Sinha and Tomaso Poggio
critical if we hope to achieve a comprehensive un- early areas or do they merely modify the conscious
derstanding of how information flows and is pro- percept that may have a later genesis?
cessed across the di¤erent stages in the brain—is
nevertheless still quite out of reach. Several open 15.6.2 How Can the Visual System Recognize
questions remain. We end this chapter by consider- Objects in Images without Relying on Sophisticated
ing two that we believe are the most important. Perceptual Constructs?
15.6.1 What Are the Neural Mechanisms by Which In a high-level, learning-based scheme of visual
High-Level Learning Influences Early Perception? processing, the first step is recognition. Without this
bootstrapping step, relevant object- and class-specific
Clearly, corticocortical feedback projections are ob- knowledge cannot be brought to bear on a percep-
vious candidates for such mechanisms (Mumford tual task. The challenge of devising a recognition
1992; Ullman 1995), although there is only a small scheme here is made more di‰cult by the require-
body of experimental data implicating the projec- ment that the strategy not rely on any sophisticated
tions in this role. Hupé et al. (1998) have found area perceptual constructs such as three-dimensional
V5 lesions to reduce responses in areas V1 and V2, shape estimates of objects in the image (because that
presumably due to disruption of cortical feedback. is what the high-level, learning-based scheme is sup-
Roelfsema, Lamme, and Spekreijse (1998) have posed to yield as its output). We have an ongoing
found heightened edge responses to the attended e¤ort to explore ‘‘minimalist’’ object representation
stimuli, perhaps due to high-level modulatory influ- schemes constructed from elementary image mea-
ences. Zipser, Lamme, and Schiller (1996) have surements (Sinha 1994, 1995; Thorek and Sinha
reported contextual modulation of responses of area 2001). Although still in an early stage of develop-
V1 cells, although it is not clear whether the modu- ment, the results thus far have been very encourag-
lation is due to lateral connections in area V1 itself or ing (figure 15.25). Other fascinating open issues that
top-down influences from extrastriate areas, and if deserve close research attention in future models of
the latter, which extrastriate areas and whether the top-down processing include methods for initial ac-
influence is object specific. quisition of high-level knowledge from bottom-up
Demonstrating, as the aforementioned studies do, measurements and schemes for combining bottom-
that the feedback connections have some modu- up data with top-down expectations.
latory influences is only part of the story. What is The prospect of tying together early perception
important, and thus far unknown, is whether these and high-level learning is an extremely exciting one.
modulatory influences are high-level and object- It represents a qualitative change in how we view
specific influences. After all, object-specific influ- the brain’s processing hierarchy and promises to help
ences need not necessarily modify the activity of unify the e¤orts in perception and cognition, leading
early areas. They could be combined with bottom- to faster progress on both fronts.
up information and thus incorporated in perception
by areas even further along the processing pathway.
Thus still another fundamental open question is, do
high-level influences actually modify the activity of
Learning to Recognize Objects
Guy Wallis and Heinrich Bültho¤
16
Abstract We e¤ortlessly convert the flat retinal images supplied
by our eyes, into a rich three-dimensional world,
This chapter reviews a large body of literature describing filled with licorice and ladybugs, shipyards and
how experience a¤ects recognition. Results both from woods. The apparent speed and ease with which we
neurophysiology and psychophysics provide clear evidence
do this is deceptive. The images cast on our retinas
for the development of recognition over time. In particu-
lar, we explain how perceptual learning in recognition can be by objects change drastically as a function of view-
directly linked to learning in feature-tuned inferotemporal point, lighting, size, or location. Consider, for ex-
lobe neurons. We argue that all of the available evidence ample, the scene depicted in figure 16.1, in which
points to the representation of objects as groups of asso- the same o‰ce chair appears several times. We seem
ciated, two-dimensional views. Further, we argue that the
to find it trivial to distinguish cast shadows or wall
natural environment is so structured that potentially very
di¤erent images appearing in close temporal succession are paintings from the genuine article, and it seems self-
likely to be views of the same object, and that this tem- evident that the chair on the desk is small enough to
poral structure allows the visual system to associate diverse hold in the hand, whereas the chair in the adjacent
views into coherent representations of individual objects. o‰ce is large enough to sit on. We happily conclude
As an introduction to the subject we review the case of this from various cues in the image, even though the
S.B., a patient whose insusceptibility to visual illusions and
images formed on our retinas by the two chairs are
failed perception of depth all point to the fact that much of
our ability to interpret the form of objects and scenes is actually identical. This chapter describes theories of
learned. We go on to describe how, by using novel stim- how humans solve the recognition problem, and
uli, it has been possible for researchers to complement particularly, how our perception of objects changes
patient studies with controlled studies of normal popu- with experience. The question of how we recog-
lations, providing us with a more precise description of
nize objects is an active area of research, and part of
how object representation and recognition develops.
The results described in this chapter strongly support the this chapter is dedicated to a summary of the pro-
empiricist view that object recognition and categorization posals that have been made. This is followed by a
is largely an ongoing process, a¤ected by experience of review of the evidence for perceptual learning in
our environment. Taken as a whole, the results serve to object recognition, ranging from the level of single
underpin the main tenet of this book, namely, that per- neurons to that of human behavior. The chapter
ception is mediated via a dynamic learning system, the
concludes by considering how we might learn to
modification of which continues throughout our lives.
associate very dissimilar views of an object, describ-
ing how temporal as well as spatial correlations
present in our environment can be used to make
16.1 Introduction the necessary associations.
Figure 16.1
Complex scene comprising many chairs seen with di¤erent sizes, viewpoints, lighting conditions, and so on, demonstrating
the range of problems faced in recognizing and categorizing objects.
rebuke. Idealist philosophers such as Kant argued 16.2.2 The Case of S.B.
that perception requires a framework, an assumed
space and time, and a concept of categories to be Having reviewed the idealist versus empiricist debate,
able to begin to represent the real world. This argu- let us consider some of the evidence for the two
ment was later championed by the Gestalt move- philosophies. The earliest arguments in favor of per-
ment, which strongly influenced thinking in the first ceptual learning stem from the work of the early
half of the twentieth century. Gestalt psychologists empiricists. Of course, being philosophers rather than
such as Köhler and Ko¤ka took the view that human scientists in the modern sense, they relied more on
perception is littered with assumptions that are used reasoning than on experimentation. Their preferred
to transform the retinal image into an object. They approach was to present readers with a mental co-
both believed the principles of organization that they nundrum followed by an elegant explanation that
proposed to be fundamental, like laws of physics, furthered their cause. Nevertheless, through such
enforcing unavoidable and universal constraints on mind games, they made at least some seemingly test-
perception. able predictions. One favorite concerned a man born
Köhler (1947, 277) condemned the empiricist blind, who is suddenly able to see. They speculated
view, stating: on how his tactile experience of the world would
transfer to his interpretation of a visual world.
Now, when the concept of organization was first intro-
Commenting on a draft of Locke’s Essay on Hu-
duced, we were at every step hampered by empiricist ex-
planations. . . . It has been shown, I hope, that [Gestalt man Understanding in early 1693, Molyneux (Locke
laws] do not allow of explanations in terms of learning, 1708, 37–38) concluded with ‘‘a jocose prob-
and that therefore, organization must be accepted as a pri- lem: Suppose a man born blind . . . and taught by
mary phase of experience. At present we may go further his touch to distinguish between a cube and a sphere.
and claim that, on the contrary, any e¤ects which learning
Suppose then . . . the blind man made to see; query
has on subsequent experience are likely to be after-e¤ects
of previous organization. whether by his sight . . . he could now distinguish
and tell which the globe and which the cube.’’ Both
Nowadays, many of the Gestaltist laws seem rather Locke and Molyneux thought not.
vague and anecdotal, but their work did succeed in Did anyone ever ‘‘live out’’ Molyneux’s gedanken-
highlighting a large number of instances in which experiment? Several hundred years ago a man born
humans infer form and shape on the basis of a few blind was almost certain to remain so, and in more
built-in assumptions. Perhaps their only disservice to modern times operable cases are usually dealt with
science was their success, which stifled progress on soon after birth. It turns out, however, that there are
perceptual learning throughout the early part of a very few cases of people recovering their sight after
the twentieth century. It was not until the 1960s years of blindness. Gregory and Wallace (1963) re-
and 1970s that interest in the seminal work of late- view several such cases making reference to the dis-
nineteenth-century empiricist writers such as Helm- cussions of the empiricist philosophers, describing in
holtz and James enjoyed a resurgence of interest. detail experiments conducted with S.B., a man who
One of the tasks for current researchers is to provide lost his sight at the age of ten months and then had it
evidence for how much of visual perception is al- restored some fifty-two years later. Finally, after two
tered by experience and how much is innate. hundred years of waiting, it seemed that S.B. could
302 Guy Wallis and Heinrich Bültho¤
been normal, developmental studies have shown that a manner quite unlike that in normal subjects. This
by this age he would have gained some ability to was confirmed by Cohen et al. (1997), who used
process depth information, including the rigidity as- magnetic stimulation to disrupt processing in partic-
sumption (Gibson, Owsley, and Johnston 1978), and ular brain areas to show that blind people require
the use of stereoscopic disparity (Held 1999). None visual cortex to interpret Braille, once again, unlike
of this early experience seems to have been retained, normal subjects. Bearing this in mind, we should shy
however; instead, S.B.’s experience with his tactile away from extrapolating too much from S.B.’s ex-
world appears to have formed the basis for what he periences. Nevertheless, the case does raise some
later saw. Indeed, there is some evidence that, intriguing questions about the influence of experi-
shortly after the operation that restored his sight, his ence in interpreting our visual environment.
previous tactile experiences caused curious mis-
perceptions of objects. For example, he initially saw 16.2.3 Pragmatism in Perception
buses as having spoked wheels, as they had had at the
turn of the century when he last had cause to touch For many years, researchers (e.g., Gregory 1972;
them as an inquisitive boy. He only correctly per- Fahle, Edelman, and Poggio 1995) have been aware
ceived the modern solid wheels months after his of perceptual di¤erences arising directly from peo-
operation. On the other hand, he had no di‰culty in ple’s experience, even as adults. One example re-
correctly perceiving the shape of a quarter moon, ceiving renewed attention concerns an illusion
which as a blind man he had imagined would be described by Pollock and Chapais (1952). This illu-
sliced like quarter of a cake! sion causes subjects to overestimate the length of
Over the next few months, S.B.’s conception of vertical lines relative to horizontal ones, which Bad-
things around him continued to improve. Gregory deley (1997) explains in terms of the level of image
and Wallace (1963, 35) noted that since leaving the correlation occurring at di¤erent orientations within
hospital, S.B. had become fascinated by the varying natural scenes. Evidence supporting this hypothesis
appearance of objects: ‘‘Quite recently he had been comes from the reliable di¤erence in the magnitude
struck by how objects changed their shape when he of the horizontal-vertical line length illusion between
walked round them. He would look at a lamp post, country folk from the Norfolk Fens and townsfolk
walk around it, stand studying it from a di¤erent as- in the City of Glasgow (Ross 1990), environments
pect, and wonder why it looked di¤erent and yet the containing very di¤erent amounts of image correla-
same.’’ It is clearly tempting to read a great deal into tion at di¤erent orientations (figure 16.3). This work
S.B.’s words, but as Gregory and Wallace point out, requires further corroboration, but if correct, it pro-
this temptation is dangerous because we cannot see vides remarkable evidence that the characteristics of
directly into the mind of a blind man. A blind man’s our everyday visual environment directly a¤ect basic
vocabulary derives from that of the sighted, and one perceptual judgements such as line length.
should therefore exercise caution in interpreting his Despite the considerable evidence supporting per-
descriptions of visual experience. Indeed, a recent ceptual learning, there are many counterexamples in
brain imaging study by Sadato et al. (1996) has dem- the literature. For example, an earlier study by Ross
onstrated that the visual cortex of blind people and Woodhouse (1979) on the same population of
becomes recruited for tasks such as Braille reading in city and country dwellers found no influence of en-
304 Guy Wallis and Heinrich Bültho¤
Figure 16.3
Images of Glasgow (left) and the Somerset Levels (right). Like the Norfolk Fens, the Somerset Levels were large, flat flood-
lands that have been drained for farming. Glasgow, in contrast, is a large industrial city littered with tall, closely packed
buildings. Daily exposure to vertical structures typical of cities may be responsible for the reduced size of the classical hori-
zontal line length illusion in city dwellers compared with country folk, as reported by Ross (1990). Such results provide
evidence that our visual diet a¤ects fundamental perceptual judgements. (Pictures ( 1997 Martin Smith, and ( 1998 Pete
Harlow, reproduced with permission.)
vironment for sensitivity to di¤erences in line ori- some systems do not adapt, arguing, for example,
entation. Also, in the field of depth perception, it has that the impression of distance derived from appar-
been shown that we perceive depth from cast shad- ent object motion during head movements (motion
ows on an built-in assumption that the light source is parallax) is not adaptable, and that it forms an anchor
situated above and to the left of the viewed object for adapting other sources of information such as
(Ramachandran 1988). Preliminary evidence that stereoscopic vision. On the other hand, recent work
this assumption is innate was provided by Hersh- has revealed a great deal of evidence that all cues to
berger (1970), who showed that without prior ex- depth can be overridden in the presence of strongly
perience of shadows, chicks perceive depth from cast competing cues (Landy et al. 1995), and that the cues
shadows. interact in a nonlinear manner (Bradshaw and Rog-
Apart from innate assumptions for interpreting our ers 1996). Indeed Wallis and Bültho¤ (1998) have
environment, some psychologists have also claimed shown that in the presence of additional depth cues,
that certain cues or sensory modalities are immutable stereoscopic information recalibrates perceived depth
in the presence of conflicting evidence, forming the from motion, directly countering the argument made
framework within which other cues are adapted. For by Spelke.
example, Harris (1963) showed that the relearning The point in raising these examples is to make it
of hand-eye coordination when wearing eyeglasses clear that neither the idealist nor empiricist view is
with displacing prisms is purely motor based, rather exclusively correct. It is important to bear in mind
than vision based. Spelke (1990), too, suggests that that the force that has shaped us, evolution, is both
Learning to Recognize Objects 305
eclectic and pragmatic. If some advantage is to be 3-D features was, for Marr, achieved in three con-
had from hardwiring certain assumptions, while secutive stages: first, the primal sketch, a wholly
leaving others to be discovered, then there is nothing two-dimensional representation which contains in-
to say that evolution has not devised such a com- formation about lines and edges visible in the scene;
promise. It is also possible that some of the basic second, a 2 12 -D sketch derived from these edges and
assumptions shared with animals as distant from us as depth information, describing closed surfaces in
chickens, are useful leftovers from before the rapid space; and finally, a full, three-dimensional repre-
expansion of the neocortex. As Ramachandran sentation of our environment built from the identi-
(1985) puts it, we should not be surprised to find that fied surfaces.
evolution has supplied us with a ‘‘bag of tricks’’ for The third stage, Marr argued, provided all of the
interpreting the world. Ultimately, what is interesting information required to recognize objects. Recog-
for those investigating perceptual learning is how nition itself involved taking the shapes drawn from
many of these tricks are inherited assumptions and the environment and matching them to stored three-
how many are extrapolated from our environment. dimensional models. These models were themselves
For that reason, recent perceptual models have built up of constituent parts, or building blocks,
focused on the use of Bayesian mechanics, in which which defined an object’s shape at various levels of
assumptions can be formally incorporated as statisti- abstraction and detail. Within such a scheme, the
cal priors (Bültho¤ and Yuille 1996). form of a standing human, say, could be said to fill
In the rest of this chapter we shall concentrate on the volume of an upright cylinder, as could a tree or
the mechanisms underlying the representation and a high-rise, but not a car or a bed, whose major axis
recognition of objects, attempting to get to the core is horizontal. Beyond this we recognize the six major
of why, as S.B. put it, an object can ‘‘look di¤erent bodily divisions of the head, trunk, and four limbs,
and yet be the same.’’ In particular, we shall explain which a¤ord discrimination from most nonanimal
why we believe that the theories of the early empiri- object categories. This recursive analysis then pro-
cists have relevance to the problem by detailing the ceeds to the level required to solve a particular dis-
role of experience both in representing and recog- crimination task or make a specific categorization
nizing objects. judgment.
Although strongly associated with this type of
hierarchical approach, Marr was neither the only nor
16.3 Object Recognition Paradigms
the first person to propose using it in categorization
and recognition. Indeed, the idea underpins a whole
16.3.1 Extracting Three-Dimensional Information
series of theories (Guzman 1971; Marr and Nishihara
1978; Brooks 1981; Tversky and Hemenway 1984)
One of the most influential writers in the field of
that can be traced back to early attempts to build
object recognition was David Marr (1982), who
artificially intelligent systems in the 1970s. Irrespec-
believed that recognition of an object requires the
tive of the detailed implementation, the approaches
matching of elemental parts of that object to the
are united in the assumption that we use 3-D infor-
parts of three-dimensional models which we have
mation from our environment to extract 3-D parts,
memorized. The correct apprehension of those
and that objects are represented as configurations of
306 Guy Wallis and Heinrich Bültho¤
Figure 16.4
When we view a surface at some random orientation in space, it often results in a characteristic two-dimensional pattern on
our retina from which we can infer its true shape. Although triangles remain triangles and ellipses remain ellipses irrespective
of viewpoint, we must decide which of the many possible shapes is responsible for the image seen. In the above examples,
the image I is due to object B, but it could equally well have been due to object A or object C.
these parts. One can think of it as a ‘‘LEGO repre- and that ellipses are views of ellipses of equal or
sentation.’’ The only major di¤erence in each case is smaller aspect ratio.
the precise shape and range of LEGO bricks used. There are several di¤erent levels of tolerance to
Examples include: polyhedrons (Waltz 1975), spheres projective distortion that the human visual system
(Badler and Bajcsy 1978), cylinders (Nevatia and might exhibit. At one extreme, there is full pro-
Binford 1977; Marr and Nishihara 1978), and ‘‘sup- jective invariance (Duda and Hart 1973; Cutting
erquadrics’’ (Pentland 1986). 1986; Weiss 1988), which assumes that full three-
dimensional information can be recovered from the
16.3.2 Projective Invariants two-dimensional image on our retina, although pro-
jective transforms can leave objects unrecognizable
A quite di¤erent approach to identifying objects is —suggesting that humans cannot achieve this.
the use of projective invariance. Projective invari- Alternatively, humans might simply ignore the e¤ect
ance refers to the fact that projection of a three- linear perspective has on the appearance of objects,
dimensional shape onto a flat surface (like our retina) namely the narrowing of straight lines with distance
produces certain characteristic patterns irrespective (Ho¤man 1966; Lamdan, Schwartz, and Wolfson
of the angle at which that 3-D feature is being 1988; Koenderink and van Doorn 1991). Unfortu-
viewed. For example, because a triangle remains a nately, this type of ‘‘a‰ne’’ approximation cannot
triangle from all but the most contrived viewing distinguish simple shapes such as rectangles because
directions, if we detect any surface with three sides, they are all a‰ne transforms of each other.
we can label it as a triangle (figure 16.4; cf. chapter The third and most promising type of invariance
15). We can then use the presence of the triangle as to be investigated is perspective invariance (Mundy
the basis for working out what the object is that and Zisserman 1992; Pizlo 1994). Perspective invari-
includes this feature. Other useful invariances are ance relies upon the types of predictable mappings
that parallel lines suggest parallel lines in the object of triangles, circles, and the like mentioned above,
Learning to Recognize Objects 307
vary considerably, but in practice all matching ap- Results consistently point to a decrease in recognition
proaches are one of two conceptually important performance as a function of the viewpoint’s disparity
types. The first uses stored models containing ex- from a previously learned view (Shepard and Cooper
plicit three-dimensional shape information, and 1982; Rock and DiVita 1987; Bültho¤ and Edelman
therefore assumes that it is possible to extract the 1992; Tarr and Pinker 1989; Jolicoeur 1990). Similar
location of three (or more) anchor points in 3-D drops in recognition performance with viewing angle
space, which are matched to those in the stored have also been reported for unfamiliar faces (Troje
models. Matching the anchor points requires a 3-D and Bültho¤ 1996).
rotation and scaling of the stored model until the These results have led to a new alternative for
anchor points are most closely aligned. Recognition how objects are represented and recognized, namely,
then proceeds by measuring the amount of overlap the feature-based, multiple-view approach (Bültho¤
in the two views (e.g., Ullman 1979). and Edelman 1992). Although it bears some relation
The second approach relies on representations to earlier two-dimensional matching theories and
based upon groups of two-dimensional views. For similarly benefits from the result of Ullman and
example, in elastic pattern matching a nonlinear Basri (1991) that any 2-D projection of a three-
image transformation is made to the incoming image dimensional object can be written as a linear combi-
of the object being viewed. A measure of how well nation of 2-D views, the multiple-view model di¤ers
the model matches the stimulus is derived by attrib- from that of the classical 2-D models in two impor-
uting a cost to how far points in one image have to tant respects. First, the views are not deformed to
be moved to find a similar-looking feature in the match each incoming image, and second, the views
other. Features which have been tried include: are represented, not as single templates, but as a col-
Gabor-like patches or ‘‘jets’’ (Buhmann, Lades, and lection of small picture elements, each tolerant to
von der Malsburg 1990), specific features such as small view changes. In the feature-based scheme,
end-stopped lines or junctions (Hinton, Williams, individual neurons are selective to features that occur
and Revow 1992), and edge-based facial features frequently in the environment. Although these fea-
such as ovals for eyes and a triangle for a nose (Yuille tures may be selective for identifiable things such as
1991). noses or eyes, most will be responsive to more ab-
stract combinations of edges and surface textures. An
16.3.5 Recognition Based on Two-Dimensional ensemble of many hundreds of cells would then be
Image Features required to act in unison to uniquely identify any
one object. The emergent properties of robustness to
Although the recognition of familiar, everyday ob- small variations in the input image (from changes in
jects proceeds almost e¤ortlessly, some views are view, size, or location) and to cell damage, have long
generally easier to recognize than others, both in been realized by the neural network community
terms of reaction times and accuracy. Such views are (Hinton, McClelland, and Rumelhart 1986).
referred to as ‘‘canonical’’ in the recognition literature Unlike other two-dimensional representation
(Palmer, Rosch, and Chase 1981). Many researchers schemes, the feature-based, multiple-view approach
have since studied view specificity using novel objects represents a significant departure from object-based
trained in particular views (see figure 16.5, panel b). models because it requires neither the extraction of
Learning to Recognize Objects 309
depth information nor the exhaustive matching of learning in inferotemporal cortex has now been
three-dimensional models. It is also consistent with a reported by many researchers (Rolls et al. 1989;
great deal of neurophysiological evidence, as we shall Miyashita 1993; Logothetis and Pauls 1995; Koba-
describe in section 16.4. (For a more detailed dis- take, Tanaka, and Wang 1998; see chapter 6). The
cussion of the pros and cons of this and the other link from view-based recognition to representations
representation schemes, see Wallis and Bültho¤ in inferotemporal cortex was strengthened through
1999; Pizlo 1994; Tarr and Bültho¤ 1995; Bieder- recording work by Logothetis and colleagues
man and Gerhardstein 1993.) (Logothetis and Pauls 1995; Logothetis, Pauls, and
Poggio 1995), in which monkeys were trained to
recognize particular aspects of the paper clip stimuli
16.4 Learning from Examples
originally used by Bültho¤ and Edelman (1992) (see
figure 16.5, panel b). After training, many neurons
16.4.1 Neurophysiology
were shown to have learned representations of par-
ticular paper clips, including some selective to spe-
From lesion studies and cellular recording it has been
cific views. In addition to longer-term changes to
proposed that a series of cortical regions starting in
cell selectivity, there is also good evidence of almost
area V1 and running ventrally through the occipital
instantaneous learning in IT cells. Tovee, Rolls, and
into the temporal lobe (V1-V2-V4-intraparietal
Ramachandran (1996), for example, presented images
areas) solves the problem of what we are looking at.
of strongly lit, two-tone (black-and-white) faces,
In contrast, a second stream leading dorsally and into
referred to as ‘‘Mooney faces’’ in the literature (figure
the parietal lobe (V1-V2-V3-intraparietal areas) has 16.7, panel a). Some IT neurons that did not respond
been implicated in the role of deciding where that
to any of the Mooney faces did so if once exposed to
object is (Farah 1990; Ungerleider and Mishkin the standard gray-level version of the face (see panel
1982; Goodale and Milner 1992; Young 1992; see
b), which accords with findings in humans, who
figure 16.6).
often struggle to interpret Mooney face images the
Cells in the latter part of the ventral stream, in the
first time, but then have no di‰culty in seeing
inferotemporal (IT) areas, are of particular relevance
them as faces a second time, even weeks later
to object recognition because of their tolerance to
(Ramachandran 1994).
changes in the precise appearance of their preferred
stimuli. Transformations tolerated by IT cells include
16.4.2 Psychophysics
changes in an object’s position, viewing angle, or
size, as well as in its overall image contrast or spatial There is considerable psychophysical evidence that
frequency content (Rolls 1992; Desimone 1991;
our perception of objects is a¤ected by experience.
Tanaka et al. 1991; see chapter 4)—indeed, they
Even a few days or hours of training can a¤ect the
include all of the types of transformation invariance
speed and accuracy with which we recognize objects.
required for view-invariant object recognition.
Bültho¤ and Edelman (1992), for example, were able
These neurons are also of interest in that they pro-
to show that if subjects learn to recognize two views
vide a source of evidence of learning in the recog-
of a novel object, recognition performance is better
nition system. Evidence for experience-dependent
for new orientations located between the two train-
310 Guy Wallis and Heinrich Bültho¤
Figure 16.6
Lateral view (upper right) and coronal section (upper left) of the primate cortex, showing some of the significant visual pro-
cessing areas. The expanded coronal section (lower left) portrays some of the important subdivisions of the temporal lobe.
(Adapted from Rolls 1992; Perrett et al. 1992.)
ing views (INTER) than outside them (EXTRA), When a view of a novel object is presented, many
which in turn is better than for orientations away feature-selective neurons respond, and the associated
from the axis linking the trained views (ORTHO; see pattern of responses they produce comes to represent
figure 16.8). A first step to explaining these results, the presence of that object. Now identification of
within a feature-based representation scheme, is to a novel view of the object will clearly be easiest for
understand why recognition performance drops with views nearest to the one trained because these views
distance from a learned view. We can start by first are most likely to contain one or more of the features
imagining what happens when leaning a single view. supporting the representation of the learned view.
Learning to Recognize Objects 311
Figure 16.7
(a) Example of a Mooney face, similar to those used by Tovee, Rolls, and Ramachandran (1996). Subjects or face-selective
neurons exposed to such a two-tone image often fail to see a face. (b) On seeing the veridical image, both neurons and
subjects can now identify the face and will continue to do so in the future, providing evidence for rapid and lasting learning.
Figure 16.8
If two views of a novel object are learned, recognition is better for new viewing angles located between the two training
views (INTER) than outside them (EXTRA), itself better than for orientations away from the axis linking the trained views
(ORTHO).
312 Guy Wallis and Heinrich Bültho¤
Using a similar line of argument, if two views of that su‰cient exposure to a particular stimulus type
the object have been learned, and both are identified causes the representation of these stimuli to alter and
as being the same object, then the presence of any of be enhanced. This finding in turn relates to the those
the features seen in either learned view will tend to of researchers studying learning in IT neurons
evoke recognition of that same object. This second (Miyashita 1988; Logothetis and Pauls 1995; Koba-
step is important because the INTER and EXTRA take, Tanaka, and Wang 1998; see chapters 4, 6),
results follow as a natural consequence, because any whose work showed that extensive experience of a
view falling within the range of the two trained class of images or objects causes a rise in the amount
views is more likely to have features in common of cells selective for those stimuli. By devoting more
with either or both of the trained views than a view neural hardware to the representation of the features
of the object from outside that range, INTER views present in an object class, one would presumably be
are more likely to be easily recognized than EXTRA better able to discriminate subtleties in their form, as
views. suggested by the Schyns studies.
The ORTHO e¤ect stems from the type of train- Gauthier and Tarr (1998) also make this point.
ing used. The training views used were not stationary Their study revealed that experience of an originally
but rather rocked back and forth through a few novel object class heightens the subjects’ awareness
degrees, which had the e¤ect that only cells tolerant of small changes to objects within the same class, and
to changes in the object’s appearance along the they argue that our highly sophisticated ability to
training meridian (see figure 16.8) were strongly recognise faces is simply due to a natural concentra-
activated during learning, and hence only they tion of neural resources from our lengthy exposure to
strongly supported recognition of the object. Be- the particular object class we call ‘‘faces’’ (see chapter
cause views of the object lying along the training 17). Some researchers, drawing on studies of the
meridian (i.e., INTER and EXTRA views) are much neurological disorder prosopagnosia, would counter
more likely to contain the features for which these that face recognition is special. Prosopagnosia is
cells are selective than views lying on an orthogonal characterized by normal ability to recognize com-
meridian (ORTHO views), INTER and EXTRA mon objects coupled with extreme di‰culty in rec-
views are more readily recognized. ognizing people’s faces (De Renzi 1997). That the
Edelman and Bültho¤ (1992) investigated the locus of the brain damage in prosopagnosia patients
e¤ects of extensive training to see whether it could was a part of the temporal lobe homologous to that
override view specificity. After training large num- of cells selective for faces in monkeys (Rolls 1992;
bers of views, they were able to change the shape of Desimone 1991) made a strong case for the sugges-
the recognition curves. Not only did reaction times tion that prosopagnosia was caused by damage to
decrease and accuracy increase, but view-specific these cells (Farah 1990). Although psychological
e¤ects, such as canonicality, gradually disappeared. studies (Tanaka and Farah 1993; Fiser, Biederman,
This issue has been raised again recently in several and Cooper 1996) have revealed a dissociation be-
studies investigating how continued exposure to an tween face and object recognition in the past, the
object class may a¤ect the manner in which the latest picture from the neurophysiological evidence
objects within the class are represented. Schyns, is not as clear as some theorists had first hoped.
Goldstone, and Tribaut 1998 and Schyns 1998 argue Direct attempts to find the illusive area responsible
Learning to Recognize Objects 313
for face recognition in monkeys has been contro- then be fed into a final decision unit. As it turns out,
versial and until now unfruitful (Perrett et al. 1992; the use of prenormalization is contrary to the evi-
Cowey 1992), which in turns lends greater weight to dence we have from the responses of real neurons
Gauthier and Tarr’s proposal (1998) that proso- implicated in object recognition. Invariance seems
pagnosia may reveal a general deficit in the area to be established over a series of processing stages,
dedicated to fine-level discriminations of highly starting from neurons with restricted receptive fields
trained objects, rather than to a specialist face area and culminating in the types of cell responses found
per se. in inferotemporal cortex mentioned earlier. With this
Apart from questions of recognition speed and in mind, it remains to be explained how one might
accuracy, there is also the question of familiarity, learn to associate very di¤erent views of an object.
which by definition is experienced based. Never- One possible solution to this problem is that in the
theless, one interesting prediction to come out of the real world, we tend to see discrete sequences of
feature-based approach to recognition is that a face images of an object, images that often undergo trans-
made up of previously experienced features, though formations. Regularity in time may act as an impor-
itself novel to the observer, should appear familiar. tant cue for predicting the identity of an object as it
This hypothesis has been tested by Solso and undergoes transformations, due to a change of posi-
McCarthy (1981), whose subjects were presented tion relative to the object. This change in viewing
with photo-fit pictures of people and then tested on position may be simply due to our approaching the
a familiarity task. The test set of faces contained ei- object, watching it move, rotating it in our hand,
ther familiar faces, wholly novel faces, or novel faces and so on. If the time domain is truly influential
containing combinations of features present in the in setting up representations of objects, then there
familiar ones. The most intriguing result was that should presumably be some evidence for this in the
subjects chose the composite faces as more familiar learning of inferotemporal neurons. In e¤ect, one
than both the unfamiliar and the familiar faces. This should expect to see quite di¤erent views of an ob-
result not only provides further support for the ject being associated to the same neuron in prefer-
distributed-feature-based approach; it also demon- ence to other very similar images, simply on the basis
strates that perceived familiarity need not correlate of the sequence in which they are presented. This
with true familiarity. last section discusses evidence that temporal relations
in the appearance of object views do indeed a¤ect
16.4.3 Temporal Continuity as a Cue to learning.
Invariance Learning The temporal association hypothesis has been dis-
cussed in the past, and has been successfully used
Although a broadly tuned feature-based system of in various neural network models of recognition
the type advocated in this chapter would be su‰cient (Edelman and Weinshall 1991; Földiák 1991; Wallis
to perform recognition over small transformations and Rolls 1997). In particular, Wallis and Baddeley
(Poggio and Edelman 1990), associating images over (1997) demonstrated how the temporal statistics of
larger shape transformations would require either the real world can be optimally used to establish
separate prenormalization for size and translation of transform-invariant representations of objects. The
the image or separate feature detectors, which would hypothesis has also found direct experimental sup-
314 Guy Wallis and Heinrich Bültho¤
Figure 16.9
Example of a pair of faces used in Wallis and Bültho¤ (2001). Each association sequence consisted of the two faces (a and b )
in profile and frontal view, and a morphed face (ab ) shown at G45 . Subjects saw both sequences (a) and (b) during training.
Figure 16.10
Results from Wallis and Bültho¤ (2001). Subjects were asked to discriminate faces previously seen in morphed sequences. (a)
Discrimination performance across training blocks, measured as percentage correct in the mismatch trials, showing for
comparison within-group (WG) and between-groups (BG) scores. (b) Same results broken down across the two training
days. ¼ p < 0:1, ¼ p < 0:01, ¼ p < 0:001.
Learning to Recognize Objects 315
port from neurophysiological recordings (Stryker clearly demonstrate that discrimination performance
1991; Miyashita 1993). Miyashita (1988), for exam- was indeed worse for faces associated in sequences.
ple, was able to show that repeating a temporal se- The di¤erence between the WG and BG condition
quence of randomly selected fractal images establishes is also seen to increase with each session of training.
cells in inferotemporal cortex that respond to one Also looking at the influence of temporal order
stimulus in the series very strongly, but also to those on the representation of objects, Stone (1998) used
patterns appearing in close succession. He was also amoeba-like shapes, similar to those in Edelman and
able to show that the e‰cacy of a stimulus declined Bültho¤ (1992), rather than wire-frame or familiar
purely as a function of temporal disparity between facial objects. During a learning phase, subjects had
stimuli. to discriminate four objects from numerous dis-
Until recently, there was little or no psychophys- tractors. In this first phase, all stimuli rotated in one
ical evidence to support the theoretical and neuro- particular direction. During testing, certain of the
physiological findings. However, in 1996 Sinha and trained objects were rotated in the opposite direc-
Poggio (see chapter 15) used temporal sequences to tion, which caused a fallo¤ in discrimination perfor-
establish the perception of the form of ambiguous mance, and an increase in reaction times. In fact,
wire-frame objects, and Wallis and Bültho¤ (2001) although similar to the results described above, the
considered the e¤ect of temporal sequences for nat- Stone (1998) results suggest something new, that
ural objects such as faces. Wallis and Bültho¤ (2001) temporal information forms part of the representa-
hypothesized that exposing observers to sequences tion of the object.
of di¤erent faces would cause them to confuse the A time-based association mechanism that correctly
identity of faces seen together in a sequence, which associated arbitrary views of objects without an
should then become apparent by the increased explicit external training signal could overcome
number of discrimination errors for faces that were many of the weaknesses of using supervised training
seen in sequences, as opposed to faces that were not. schemes or of associating views simply on the basis of
Figure 16.9 puts this hypothesis in a more graphical physical appearance. For this reason, the three ex-
light by displaying two possible sequences, each periments described above may well represent a sig-
containing two di¤erent people’s faces. The tempo- nificant new step in establishing the two-dimensional
ral association hypothesis predicts a higher confusion multiple-view approach to object recognition.
rate for pairs of faces associated in this way than for
pairs of faces coming from two di¤erent sequences.
During the experiment, subjects were exposed to
thirty-six such pairings of heads and then tested on
their ability to discriminate them. The results of their
experiment are displayed in figure 16.10. ‘‘WG’’
indicates that the tested faces were from within a
group, that is, appeared in a training sequence to-
gether; ‘‘BG,’’ that the faces tested were once again
familiar but came from separate groups, thus had not
been seen in the same training sequence. The results
This page intentionally left blank
Learning New Faces
Vicki Bruce and Mike Burton
17
Abstract ceptual categories based on subtle di¤erences between
items, unnoticed by the untutored eye. Moreover, all
As children, we become particularly sensitive to the con- of us, whether we also become accomplished radio-
figuration of upright faces and to variations that discrimi- graphers, bird-watchers, or car bu¤s, will continue
nate faces within our own race, although we have di‰culty
throughout our lifetimes to meet new people in the
recognizing faces when they are presented upside down or
show variations with which we are unfamiliar. Throughout flesh, in films, or on TV. Through these encounters,
our lives we continue to learn to recognize each specific each of us will continue to learn new members of
new face we encounter. During learning the visual repre- the most important social category of all—the human
sentation shifts away from external face features and toward face.
internal ones, particularly the eyes, and allows us to recog-
Familiar face recognition represents a considerable
nize a familiar (learned) face across a range of viewpoints,
lighting, and expressions. After reviewing what is known challenge for the human recognition system because
about how we learn faces—both generally and specifi- known faces must be recognised across a variety of
cally—this chapter outlines a theory that links the learning transformations. Like other objects, faces must be
of new faces to the mechanisms responsible for repetition recognized across di¤erent viewing angles and in
priming of familiar items. di¤erent lighting conditions. Unlike most other ob-
jects, however, faces also deform nonrigidly during
expression and speech; slower, age-related changes
17.1 Introduction in weight, wrinkles, hairstyle, and color, and some-
times rapid changes in hairstyle, also pose major
The human brain is a perceptual learning machine challenges to the recognition system.
par excellence. Di¤erent demands are made of this Early consideration of the processes of face rec-
machine at di¤erent times in its life. Developing ognition produced some apparently extraordinary
infants quickly learn to di¤erentiate and to individ- demonstrations of ‘‘one-trial’’ face learning in ex-
ually recognize the important object categories in periments reported in the late 1960s and 1970s. A
their world. At about the time children become number of researchers (e.g., Shepard 1967; Standing
familiar with most of the basic object categories they 1973) showed that recognition memory for once-
will ever encounter, they meet the new visual per- viewed pictures was extremely accurate and that this
ceptual challenge of learning to read. accuracy was maintained when pictures were drawn
The further task of distinguishing between objects from homogeneous meaningful sets of items such as
of the same overall type is one that will continue to faces or pictures of houses. Memory for faces, how-
tax the perceptual learning system throughout ever, generally exceeded that for other comparable
adulthood: almost every chosen profession, trade, or groups of items. For example, Yin (1969) asked
hobby requires its practitioners to acquire new per- subjects to study for three seconds each image from a
318 Vicki Bruce and Mike Burton
series of forty pictures drawn from two di¤erent picture-specific performance with unfamiliar faces
groups (e.g., faces and houses). An immediate forced- somehow transforms into the kind of visual memory
choice recognition test followed in which subjects that allows us to recognize highly familiar faces in
were asked to choose the studied item from each of di‰cult and novel circumstances.
twelve pairs of pictures of each type of object. The This chapter will review what is known about the
faces were all clean-shaven, adult males chosen to learning process which is relevant to the nature and
‘‘be similar with respect to general age, expression acquisition of representations of familiar faces, and
and lack of outstanding distinguishing features . . .’’ will not consider other important and interesting
( p. 142). The houses were less uniform, to Yin’s eye, questions about memory for pictures, memory for
than the faces. Nonetheless, recognition memory for relatively unfamiliar faces (as in eyewitnessing), or
upright faces was 93% correct, compared with 81% computer models of human face recognition, except
for pictures of houses. where these topics illuminate the main thrust of the
Similar high rates of recognition memory have chapter. We start by considering the developmental
been found using ‘‘old-new’’ rather than forced- learning process by which children gradually im-
choice recognition memory tests, where decisions prove in face recognition abilities until they reach
must be made to each individual face in a series of adult levels of expertise. The expertise acquired seems
previously studied and new, distractor items. For to include a particular sensitivity to the configural or
example, Bruce (1982, experiment 1) presented a relational properties of face patterns in general, plus a
series of 24 unfamiliar male faces for study, for 8 knowledge of the discriminating dimensions of the
seconds each, and later asked participants to decide specific types of faces encountered during develop-
whether each of a series of 48 male faces was old or ment. Despite this expertise, however, adults must
new. When the ‘‘old’’ faces were shown in identical learn each individual face they encounter. Briefly
pictures to those studied, recognition rates of 90% exposed, thus poorly learned faces are extremely
were recorded, with a false positive rate of 12%. di‰cult to recognize; the process of learning an in-
Importantly, however, in this study recognition rates dividual face seems to shift its representation from
for faces shown with changed viewpoint or expres- one of extreme image specificity to a consolidated
sion were considerably less accurate. Thus what the representation of a familiar face that can allow rec-
early recognition memory experiments usually dem- ognition across a range of natural transformations.
onstrated was very good memory for specific pictures This process of learning new faces will form the
of faces, rather than any rapid learning of representa- major focus of this chapter.
tions useful for generalising to novel views.
Indeed, in a meta-analysis of 128 eyewitness
17.2 Comparisons between Face and Object
and facial identification studies, Shapiro and Penrod
Recognition
(1986) found that one important metafactor a¤ecting
recognition memory for previously unfamiliar faces
Before considering the detail of how we learn faces,
is the extent to which the test conditions re-create
it is interesting to inquire whether there is any di¤er-
the study conditions—in terms of facial image char-
ence between learning to recognize faces and learn-
acteristics and associated context. Initially highly
ing to recognize other categories of objects. To what
Learning New Faces 319
extent do the principles of learning faces illuminate successfully identified, all faces serve a variety of other
(or become illuminated by) those applying to other functions that involve other mappings between visual
categories? (For a fuller comparison of face and ob- form and meaning. We recognize emotions from
ject recognition, see Bruce and Humphreys 1994.) facial expressions and can also make phonemic deci-
sions based on perceived mouth shapes. We make
17.2.1 Multiple Levels of Categorization use of head and eye gaze direction to understand
where others are directing their attention. These
Faces, like other objects, can be categorized at a social messages are derived from mapping between
number of di¤erent levels. First there is the ‘‘basic- di¤erent nonrigid (expressive, speaking) movements
level’’ categorization (Rosch et al. 1976; Biederman and rigid (head direction) variations and other
1987): this pattern is a face, as opposed to some other meaning categories. Such variations in the facial con-
kind of thing, such as a dog or a car. The system re- formation are likely to make the task of face cate-
sponsible for object recognition may deliver this gorization and particularly face identification di‰cult
categorization of a face as a face—or it may be part because the range of rigid and nonrigid transforma-
of a specific hardwired face detection module, which tions of the face provides essential input to other
may be crucially important for infant-mother inter- social processes. Owing to the subtleties of the social
actions and for social interactions in adults (see, for demands of face processing, the representational
example, Goren, Sarty, and Wu 1975; Johnson et al. requirements of recognition are likely to be di¤erent
1991). for faces compared with most other objects, although
Another level of categorization is of the face as a other varieties of expert identification, categorization,
kind of face—for example, as a male face, as a Japa- or both may well resemble face processing.
nese face, or even as a nice-looking face. This requires
that we derive meaning from the visual character- 17.2.3 Expertise with Faces
istics of the face alone (what Bruce and Young 1986
termed visually derived semantics), whether familiar or Born with the ability to track facelike patterns with
unfamiliar. This process is somewhat like recogniz- their eyes and head immediately after birth (Goren,
ing what breed of dog or what make of car is pre- Sarty, and Wu 1975; Johnson et al. 1991), babies
sent. Finally, we are able to determine, for familiar very rapidly learn at least some faces: within the first
faces, to whom the face belongs. This process is few days of life, they will look more at the face of
more like recognizing an individual dog or our own their mother than at another woman of similar ap-
car. It is this final level of face identification that is pearance in conditions where olfactory cues are
the most visually demanding because di¤erences be- masked (Bushnell, Sai, and Mullin 1989). Despite
tween individual faces can be extremely modest and these early proficiencies, however, face processing in
can defy description. early childhood is considerably less accurate than in
adulthood. Chung and Thomson (1995) review the
17.2.2 Other Uses of Information from Faces development of face recognition skills through
childhood, and a recent study of our own (Bruce et
Besides serving as unfamiliar objects to be catego- al. 2000) shows how a range of face-processing skills,
rized in particular ways or as familiar objects to be including the ability to process expression, facial
320 Vicki Bruce and Mike Burton
well, and Hay (1987). At all ages tested, the halves of In this view, learned experience with faces creates
aligned composites were more di‰cult to identify expertise at encoding the spatial relationships that
than those of nonaligned composites when displays distinguish between di¤erent members of the class
were upright, but there was no di¤erence between ‘‘face.’’ Diamond and Carey (1986) distinguished
aligned and nonaligned versions when displays were between first-order spatial relationships, which de-
inverted. Thus, to the extent that ‘‘configural’’ pro- fine a basic-level category such as face (eyes above
cessing is tapped by the composite e¤ect, it seems well nose above mouth), and second-order spatial rela-
established in early childhood. Nonetheless, in these tionships, which di¤erentiate members of this class.
same experiments, Carey and Diamond found an They reasoned that if expertise in distinguishing faces
age-orientation interaction—with adults showing resulted from increasing sensitivity to these configural
stronger overall decrements from the inversion of relationships, then perhaps experts in domains in-
faces than children. volving other classes of objects having the same first-
Carey and Diamond (1994) explain their appar- order representation would also show heightened
ently paradoxical findings by distinguishing between sensitivity to spatial relationships. If so, such experts
the access of parts of faces and the recognition of should show inversion e¤ects like those found with
faces on the basis of relational features. They argue faces. Diamond and Carey (1986) tested this hy-
that young children, like adults, clearly find it di‰- pothesis with dog experts ( judges of specific breeds
cult to access parts of faces from within whole ones of dog) and showed that such individuals su¤ered as
(aligned composites); hence the composite e¤ect is much in terms of decreased recognition rates from
evident at all age groups. Nonetheless, experience inverting pictures of dogs as pictures of faces. Non-
with faces through childhood enhances the sensitivity experts showed greater e¤ects of inverting pictures
to configural relationships in the representations used of faces than dogs. Although there are some prob-
to identify faces: adults are much more disadvantaged lems with the results of this study (the dog experts
when faces are inverted, and their configural rela- were not actually superior to the nonexperts on up-
tionships thus become more di‰cult to code. To use right pictures of dogs, which Diamond and Carey
a simple analogy, the young child may have a very attributed to the experts’ being much older than the
unsophisticated representation of familiar faces, for control subjects), they nonetheless support the gen-
example, that one face has big eyes and another has eral theory that expertise creates sensitivity to dis-
small ones, but the processing of the whole config- crimination of second-order spatial relationships.
uration of an aligned composite could still interfere
with the process of separately accessing the eyes to 17.2.4 Cross-Race E¤ect
determine their size. The adult may have a much
more elaborated representation of the relative shapes Although adults are expert with faces of the type
and separation of the eyes and their relationship with they have learned through exposure, they are much
other facial features. Alignment of composites will worse at recognizing faces of unfamiliar types than
make such descriptions equally hard to recover, but younger people are. Europeans find Japanese faces
inversion will additionally disrupt the derivation of di‰cult to recognize, and vice versa (see Brigham
these kinds of relationships. 1986 for a brief review). Such e¤ects cannot be ex-
322 Vicki Bruce and Mike Burton
plained in terms of inherent di‰culties of one kind 17.3.1 Unfamiliar Face Representations Are Image
of face or another, or we would not observe cross- Specific
over interactions of this kind. The best-supported
explanation is that experience with faces of our own A good illustration of the di¤erence between repre-
race allows us to learn just the features that are dis- sentations for familiar faces and those for unfamiliar
criminating within that race. In this view, increased ones comes from an ongoing project of our own
contact with members of another race should reduce on the identification of people shown from closed-
or even eliminate the cross-race e¤ect. Some evi- circuit television (CCTV) images (see Bruce 1998
dence supports this ‘‘contact’’ hypothesis, though the for an overview). In some of our studies, observers
data are rarely unambiguous. For example, Chiroro are asked to choose the person from a photo array
and Valentine (1995) found that black African who matches the person shown on video, and in
students who had a great deal of exposure to white others asked to verify whether a video image matches
faces recognized white faces as well as black ones, a photographic still for identity. Observers make
but the results obtained in the other half of the de- substantial numbers of errors identifying images of
sign with white observers were not so clear-cut. unfamiliar faces, even when the quality of the video
Thus evidence from the development of face rec- image is high, whereas they are highly accurate in
ognition and sensitivity to inversion combines with identifying and verifying images of familiar faces
that from the cross-race e¤ect to provide a reason- shown on video, their rates being at or near ceiling,
ably consistent picture. Our facility at recognizing even when video image quality is poor (Bruce et al.
upright faces arises as a result of years of exposure to 1999; Burton et al. 1999; Bruce, Henderson et al.
faces during childhood, exposure that enables us to 2001).
discriminate types of faces from their subtle config- In a recent study (Bruce, Henderson, et al. 2001),
ural variations on the basic face pattern. Zoe Henderson showed observers still images or
The recognition skills we have considered here short clips of a person walking into the University of
apply to faces in general, not just to familiar ones. Glasgow Psychology Department, where a CCTV
But what happens when a face becomes familiar? system records a short clip of each entrant, of an
How do we learn new faces? To answer these ques- image quality typical of commercial systems. They
tions, we need to understand the nature of the visual were given a still photograph of either the same
representations used when faces are unfamiliar, and person, or a di¤erent person, chosen to bear some
how these change as a face is learned. resemblance to the target person in the video clip
and asked to decide whether the two matched or
not. A group of sixty observers who were chosen to
17.3 What Is the Di¤erence between
be familiar with the people shown in the video
Unfamiliar and Familiar Faces?
images averaged 94% correct in this task, whereas
observers unfamiliar with any of the faces seen in
Although visual memory for initially unfamiliar faces
video or still images averaged only 73% correct.
appears to be dominated by details of the particular
Note that there is no memory component at all in
picture or image of the face encountered, our represen-
this task—observers must simply compare the visual
tations for familiar faces appear to be more abstract.
image of one face against another, and have unlimited
Learning New Faces 323
time to make their decision. Performance when faces change in viewpoint or lighting. Indeed, this dis-
are unfamiliar shows that the visual images of the two ruption occurs even in tasks with no memory com-
faces alone cannot be used to make an accurate com- ponent at all, where the task is simply to compare
parison of the face features. Images of an already two faces and determine whether they are the same
familiar face, however, can be independently identi- person. For example, Hill and Bruce (1996) found
fied even from the poor-quality video images, and matching accuracy was reduced when two images of
these preexisting visual categories, and the nonvisual the same person were shown in di¤erent viewpoints
(identity) information that they access, can e¤ec- or with di¤erent lighting directions. Bruce et al.
tively mediate the comparison of the two images. (1999) found that matching video images to full-face
In another of our experiments, observers were photo arrays was more di‰cult when the head pose
shown target images of unfamiliar faces extracted of the video target was changed from full face to 30 .
from high-quality video and asked which of an array Thus the image specificity of face matching does not
of ten male faces of similar overall appearance appear to be a function of memory, but of the way
matched the identity of the image at the top. Al- we perceive unfamiliar faces.
though observers knew that the target was present Familiar faces are not immune to all such e¤ects,
on every trial, they made about 20% false matches in however. Johnston, Hill and Carman (1992) found
this task (Bruce et al. 1999), even when the target that recognition of friends’ faces was disrupted if these
image was chosen to match the full-face viewpoint were shown with a highly unusual lighting direction
and neutral expression of the array faces as closely as (lit from below). Moreover, familiar face recognition,
possible. like unfamiliar face processing, is dramatically dis-
Such studies suggest that our visual representations rupted by inversion of the image in the picture plane
of unfamiliar faces are highly image specific (see also (see Valentine 1988 for a review) and by reversal
Hill and Bruce 1996; Kemp, Towell, and Pike 1997). of polarity when images are shown in photographic
Even when no memory load is involved, two di¤er- negative (see, for example, Galper 1970; Phillips
ent images of the same person can easily be confused 1972). Bruce and Langton (1994) showed that where
with images of two di¤erent but similar-looking observers were forewarned of the list of celebrities
people. Yet we do not appear regularly to make such whose faces might appear, identification of famous
confusions between people we know well, although faces dropped from 95% in upright, positive images
certain kinds of errors and di‰culties in everyday to 55% upright negative and 70% inverted positive.
person recognition are reasonably frequent (Young, The two transformations combined virtually abol-
Hay, and Ellis 1985). Clearly, there must be some ished recognition altogether, with recognition drop-
shift in the representational process as we learn faces. ping to a mere 25% correct. It seems, then, that
expertise with faces in general and familiarity with
17.3.2 Unfamiliar Face Representations Do Not individual faces lead to representations that can tol-
Allow for Generalization erate transformations within the range usually expe-
rienced (e.g., of angle and of lighting) but not outside
Recognition of unfamiliar faces is considerably it (see Bruce 1994 for further discussion).
disrupted by changed image conditions, such as a
324 Vicki Bruce and Mike Burton
17.3.3 Internal versus External Features Importantly, Young et al. (1985) found that the
shift toward internal feature processing occurred
Using recognition memory and matching tasks, Ellis, only when the identities of familiar faces had to be
Shepherd, and Davies (1979) and Young et al. (1985) matched across di¤erent pictures. When familiar
showed that representations of unfamiliar faces are faces were used in a picture-matching task, the rela-
dominated by the external features of hairstyle and tive ease of internal and external feature matching
face shape. This is consistent with similarity sort- was the same as shown with unfamiliar faces. Thus
ing studies reported by Shepherd, Davies, and Ellis it is in the more abstract ‘‘structural’’ codes, which
(1981), where the three main dimensions extracted allow the recognition of familiar faces across changes
from multidimensional scaling of similarity judg- of viewpoint, that internal features acquire their
ments corresponded to hairstyle, face shape, and age. salience.
In contrast, however, the recognition and matching
studies reported by Ellis, Shepherd, and Davies 17.3.4 Dynamic Properties
(1979) and Young et al. (1985) found that repre-
sentations for familiar faces are weighted more toward There also seem to be di¤erences in the ways that
internal features. Ellis, Shepherd, and Davis found dynamic properties of unfamiliar and familiar faces
that the external feature advantage in unfamiliar faces are used for recognition. In a study since replicated
competely switched to an internal feature advantage in our own laboratory (Lander, Christie, and Bruce
in familiar ones, whereas Young et al. found that ex- 1999), Knight and Johnston (1997) showed that
ternal and internal features were rather more equally negative images of famous faces were better recog-
weighted in familiar faces, but still showed the shift nized when an animated sequence was shown than
from external feature dominance in unfamiliar faces. when a single static image was shown. Such a result
The shift in feature salience is consistent with the might arise simply because an animated sequence
idea introduced earlier that unfamiliar face repre- conveys more information than a static image does in
sentations are based primarily on image properties of the di¤erent viewpoints and expressions shown. In
the specific pictures in which they are encountered her doctoral research at the University of Stirling,
(termed pictorial codes by Bruce and Young 1986). Karen Lander has shown convincingly that this can-
Although the hairstyle and external features repre- not be the only explanation. Using faces made hard
sent a relatively large proportion of the total image to recognize through thresholding (figure 17.2)
encountered, hairstyles and face shapes change quite rather than negation, she has found that the benefi-
substantially over time, as people restyle their hair, cial e¤ects of motion depend critically on motion
change weight, or age—much more so than do in- characteristics rather than on the number of frames
ternal features, whose characteristics change rela- shown. Thus animated image sequences are best
tively slowly with age. In contrast, internal features recognized when the face is shown at its original
are mobile and communicative and their momentary frame rate, as opposed to more quickly, more slowly,
postures must be attended to in interpersonal com- or in reverse or disrupted sequence (Lander, Christie
munication. Perhaps these are the reasons why rep- and Bruce 1999; Lander and Bruce 2000).
resentations for learned (familiar) faces become Findings from unfamiliar face recognition are
shifted to capture this area of the face more strongly. slightly di¤erent. Christie and Bruce (1998) found
Learning New Faces 325
representations for identity seem to require that the sions ( principal components) that can describe a set
overall pattern of light and dark be preserved: nega- of faces economically. The di¤erent components
tive images of faces are di‰cult to recognize because extracted can be depicted graphically as ‘‘eigenfaces’’;
they invert the usual pattern of light and dark. individual faces can then be described and recon-
Bruce and Langton (1994) and Kemp et al. (1996) structed from the sum of a set of eigenfaces suitably
report studies exploring in greater detail why nega- weighted. Thus far, we have found that the PCA
tive images are hard to recognize. Bruce and Langton coding scheme certainly seems to capture, at least in
(1994) showed that the recognition of nonpig- part, how humans encode faces, though the wavelet-
mented, three-dimensional surface images of faces based graph-matching model may be a better model
derived from scanning the shapes of faces with a laser of how we recognize faces across pictorial trans-
range finder was a¤ected rather little by negation. formations (Hancock, Bruce, and Burton 1998).
They used this result to suggest that the inversion of In sum, poor recognition of line drawings and the
brightness of pigmented regions of the face (e.g., e¤ects of photographic negation support the idea
light skin to dark skin) was responsible for the e¤ects that patterns of light and dark are critical for repre-
of negation, rather than the reversal of shading pat- senting faces. Shading patterns may lead to the deri-
terns that were preserved in 3-D surface images. vation of an explicit three-dimensional description
Kemp et al. (1996) showed that the alteration of hue of the facial surface, which is clearly important for
of pigmented regions did not a¤ect recognition of some tasks (for example, we cannot stroke someone’s
familiar faces, and suggested that the negation e¤ect face unless we understand the 3-D layout of its
arose at least in part from how negation a¤ected surface). However if 3-D shape were computed
shape-from-shading processes. explicitly as part of the representation for recogni-
Nevertheless, because even highly familiar faces tion, it is di‰cult to understand why we are so poor
are extremely di‰cult to recognize from three- at recognizing faces from new viewpoints. We sug-
dimensional shape alone, representations for face gest instead that 3-D shape is coded only implicitly
recognition seem unlikely to be based wholly or via the direct analysis of two-dimensional intensity
even largely on 3-D shape descriptions derived from patterns that mediate our recognition of identity.
shading patterns (Bruce, Healey, et al. 1991). Find-
ings such as these have encouraged our group (e.g., 17.4.2 Features versus Configuration: Holistic
Hancock, Burton, and Bruce 1996; Hancock, Bruce, Representations
and Burton 1998) to explore low-level image-based
coding schemes such as principal components anal- One of the attractions of low-level image-coding
ysis (PCA) of image pixels (e.g., Kirby and Sirovich approaches such as those mentioned in the last sec-
1990; Turk and Pentland 1991) and graph-matching tion is that these analytic techniques extract global
of Gabor wavelet patterns (e.g., Wurtz, Vorbruggen, dimensions rather than local features for describing
and von der Malsburg 1990) as possible analogies for faces. A frequent finding in recent studies of face
how the human visual system describes faces for processing has been that the configuration of face
recognition. In PCA, the intercorrelations between features is at least as important as the local features
the variations in image intensity between each pixel themselves. In section 17.2.3, we described Young,
are computed and used to derive underlying dimen- Hellawell, and Hay’s investigations (1987) of the face
Learning New Faces 327
Figure 17.3
Images illustrating the work of Leder and Bruce (1998). The central image is distorted by changing the configuration (left) or
local feature (right). Both changes are equally salient in upright faces, but only the local feature change remains more dis-
tinctive than the original when the image is inverted.
composite e¤ect, where new face identities emerged together; see figure 17.3). When such faces were
from the juxtaposition of the top half of one face with shown upright, then both types of faces were rated as
the bottom half of another, provided these were more distinctive and found to be more memorable
closely aligned. Of the many similar demonstrations than the originals. When shown upside down, how-
of the importance of configural information for face ever, only the distinctive local versions maintained
processing (e.g., Tanaka and Farah 1993; Rhodes, their advantage over the original versions. The dis-
Brake, and Atkinson 1993; Bartlett and Searcy 1993; tinctiveness of those with configural alterations com-
Searcy and Bartlett 1996), Leder and Bruce 1998 is pletely disappeared. Experiments such as these raise
notable for the realism of its configural and feature the thorny issue of what exactly is meant by ‘‘con-
manipulations. Original faces with hairstyle con- figural’’ processing of a face. Do we mean the coding
cealed were rendered more distinctive either by of the spatial relationship between face features—or
making a change to a local feature (e.g., making the is a face processed as an undi¤erentiated ‘‘whole’’
eyebrows more bushy) or by altering the relationship pattern, within which there is no discrete analysis of
between features (e.g., by moving the eyes closer features such as eyes, nose, and mouth at all? Al-
328 Vicki Bruce and Mike Burton
though many researchers are unclear about this issue, representations first become more abstracted in the
others have expressed distinct positions: Diamond process of learning individual faces. Starting with a
and Carey (1986) clearly point to the importance of representation that captures image-based properties
the spatial relationships between di¤erent face fea- of the regions and interrelationships seen in one or a
tures, and Tanaka and Farah (1993) suggest that faces very few images of an unfamiliar face, we somehow
might be processed as nondecomposed wholes. gain, over the course of many encounters with that
The PCA (or eigenface) approach to automatic face, perceptual learning of a consolidated visual
recognition is attractive in this context because of memory that now emphasizes internal over external
the global nature of the individual components (i.e. features, and allows a degree of generalization to
‘‘representational primitives’’) it uses. The issue of novel images of that face in a way not possible for
part-based versus holistic face processing arises in the unfamiliar.
part from an intuition that faces are made up of One possibility is that the di¤erent image-based
component parts, and that these di¤erent parts can encounters with the face literally become averaged
be arranged in di¤erent configurations. Because rec- in some process of overlaying or merging of succes-
ognition appears to be influenced both by the parts sive images of the face to form a prototype repre-
and by their configuration, it becomes hard to define sentation. Another is that each exemplar is stored
either independently. Using PCA as a representa- independently and that generalization emerges from
tional primitive has the advantage that faces are built the increasing opportunities for overlap between any
of sets of features (or dimensions or primitives) but new instance and one of the collection stored. Baron
each of these covers the entire face. Thus the intu- (1981) provides a clear early computer model of how
ition that there are two aspects of faces, the primitives both averaging and retention of distinct exemplars
and their configuration, may have to be replaced with could coexist. According to his model, as each in-
a representational scheme combining the two. PCA stance of a known face is encountered, its similarity
is not the only scheme available in this regard. Other to already stored representations of that face is com-
systems, such as Gabor-filter-based matching systems pared, with the fate of the new instance dependent
also rely on primitives that cover a very large part of on its degree of resemblance to what is already
the face, rather than just common-usage features stored. If its overlap with a stored exemplar is very
such as mouths or eyes. Technical solutions to engi- high, then the new instance is simply discarded,
neering problems of pattern recognition may shed whereas, if its resemblance is very low, then the new
light on the nature of the primitives in face recog- instance is stored as a novel exemplar of that person.
nition; psychological research is under way to estab- If similarity is intermediate between these two ex-
lish whether these representations may be useful in tremes, then the new exemplar is averaged with the
understanding human recognition. record already stored. In this way, for example, faces
might acquire a set of consolidated or abstracted
17.4.3 Prototype Face Representations? representations at each of a distinct number of ca-
nonical viewpoints.
Thus far, we have considered what might form the Evidence consistent with this kind of storage pro-
basis of our representational primitives for face rec- cess has been obtained in experiments on the storage
ognition, but not how picture- or image-specific of face ‘‘prototypes’’ from variant facial exemplars.
Learning New Faces 329
Bruce et al. (1991) gave observers a series of faces to 1989 for an overview). On the other hand, although
rate for apparent age. The series contained several it is easy to think about the separation of exemplars
di¤erent variants of each of a number of distinct in- within the learning of a concept such as ‘‘bird’’
dividual faces, where variations were created by dis- (where each exemplar might be one type of bird—
placing the internal features of the face upward or sparrow, robin, ostrich, etc.), it is not so easy within
downward by regular amounts around a starting or the domain of perceptual learning of categories such
‘‘prototype’’ arrangement of the features. We found as ‘‘John’s face.’’ Even though our laboratory ex-
that, in a later unexpected recognition memory test, periments may separate John’s face into a number of
observers found these ‘‘prototype’’ faces highly fa- discrete training or test exemplars ( photographs or
miliar, even if they had never been seen in the study pictures), in everyday life, a new face will be en-
series. Such findings were reminiscent of those by countered in the continuous transformations of ex-
Posner and Keele (1968) using varying dot patterns pression, gesture, and speech. In such a context, a
from which the protypical pattern arrangement was process of averaging and abstraction—at least within
apparently learned. small variations of the same viewpoint—would seem
One way of explaining these data is to suggest that more likely on logical grounds. Finally, we note that
successive instances of a face become superimposed neither of these broad approaches seems well suited
as a face is learned, so that the ‘‘average’’ of the en- to capture the dynamic properties of faces that also
countered traces seems more familiar than any indi- seem to be represented as a result of learning (see
vidual trace. Importantly, Bruce (1994) and Cabeza section 17.3.4).
et al. (1999) reported limits to this superimposition
e¤ect. When face variations were based on angular 17.4.4 Categorical E¤ects
changes in head pose, there was little evidence for a
prototype e¤ect unless angular variations were very Another shift occurs as a face is learned through the
small indeed, whereas variations in feature placement process of abstraction outlined above. The function
(internal configuration) led to prototype e¤ects, even of such learning seems to be to make di¤erences
when the feature variations shown were very large between di¤erent instances of the same face less easy
(Cabeza et al. 1998). to see because they are united by the common rep-
Thus our suggestion is that, as an initially unfamil- resentational entity that allows each of them to be
iar face is learned, di¤erent instances of the same face recognized. In contrast, di¤erences between one face
become superimposed to provide a structural code and other similar looking people become easier to
(or, most likely, a set of such codes spanning di¤er- see, and when faces are familiar they appear to be
ent viewpoints), which represents the central ten- largely immune from the kinds of confusion we find
dencies in all the variations of the face encountered. in matching images of unfamiliar faces.
Finally, we should note that within the large lit- This function of categorical perception is a familiar
erature on prototype or ‘‘typicality’’ e¤ects in per- one when applied to certain kinds of perceptual cat-
ceptual and conceptual learning there has been a egories, such as colors (Bornstein and Korda 1984),
fairly lively debate between ‘‘prototype’’ theories and phonemes (Liberman et al. 1957), and even facial
‘‘exemplar’’ theories, with the weight of evidence expressions (Etco¤ and Magee 1992; Young et al.
tending to favor the exemplar account (see Medin 1996). In all these domains, it has been shown that
330 Vicki Bruce and Mike Burton
it is harder to discriminate between two di¤erent ceptual learning, and how much from verbal labeling
exemplars that fall within the same perceptual cate- as a result of such learning.
gory (e.g., two di¤erent shades of red or two di¤er-
ent happy expressions) than between two exemplars 17.4.5 What Changes as Faces Become Familiar?
of equivalent physical dissimilarity that straddle a
category boundary (e.g., a red with an orange or a The preceding sections have suggested that repre-
happy face with a sad one). sentations for familiar faces are based upon the
Interestingly, these same kinds of e¤ects have also superposition of image-based descriptions in a way
been demonstrated in familiar face recognition. that selectively enhances their internal facial features.
Beale and Keil (1995) presented observers with Can we experimentally chart the process of acquiring
images that were morphed in graded steps between new representations of previously unfamiliar faces
two familiar faces such as Bill Clinton’s and John F. and see the changes in visual representation emerge
Kennedy’s. At one end of the continuum lay morphs in more detail? This interesting idea was tested in an
(e.g., 90% Clinton and 10% Kennedy or 80% Clin- unpublished experiment by Hadyn Ellis and col-
ton and 20% Kennedy) that were readily categorized leagues, who showed observers a set of video se-
as one or the other person (e.g., Clinton). Near the quences of a to-be-learned set of faces every day for
middle of the continuum were morphs that were several days. They found that performance on a face-
much more di‰cult to categorize (e.g., 50% Clinton matching task shifted from external feature domi-
and 50% Kennedy). They found that discrimination nance toward internal feature dominance over the
between two di¤erent Clinton-Kennedy morphs course of this training. These preliminary findings
straddling the boundary between their identities have been replicated in recent studies within our
was much better than discrimination between two own group.
morphs both judged to belong to the same identity, Adriana Angeli (see Angeli, Bruce, and Ellis 1999)
although importantly, such categorical e¤ects were attempted a similar study at the University of Stir-
not found for pairs of faces unfamiliar to participants ling: on each of nine successive days, observers were
(but see Levin and Beale 2000). introduced to short video clips of thirty individuals
Stevenage (1998) was able to chart the process of accompanied by some descriptive information about
learning familiar but di‰cult face categories using who the people were. Each day included a face
photographs of the faces of identical twins. As ob- familiarity decision test where observers had to dis-
servers became able to discriminate one twin from criminate thirty familiar from thirty unfamiliar faces
another, their similarity ratings came to exhibit cat- using whole faces, internal features, or external fea-
egorical e¤ects. After training with numerous images tures (ten familiar and ten unfamiliar items in each
of each twin’s face, they judged di¤erent exemplars condition each day). The results from this study were
of the same twin to be more similar in appearance, very promising, with a switchover from external
and di¤erent exemplars of di¤erent twins to be more feature dominance to internal feature dominance
dissimilar in appearance, than they had before training. over the course of the experiment, as shown in the
Such categorical e¤ects are pervasive, although summary of data in table 17.1.
their explanation is also controversial. It is unclear A further promising study has just been completed
how much categorical perception arises from per- by Chris O’Donnell, a doctoral candidate at the
Learning New Faces 331
Moreover, repetition priming seems to require item the architecture and simulations, see Burton, Bruce,
identities be retrieved: sex decisions do not prime and Johnston 1990.)
sex decisions, for example (Ellis, Young, and Flude This simple idea that repetition priming involves
1990). strengthening the links (or pathways) used within a
Repetition priming of faces is sensitive to varia- model has been extended to more recent versions of
tions in facial appearance that seem to reflect repre- the Burton, Bruce, and Johnston (IAC) model of the
sentations used for recognition. Priming is reduced cognitive aspects of face recognition. Burton, Bruce,
when face viewpoint, expression, or both are varied and Hancock (1999) describe how the IAC model
(Ellis et al. 1987), reduced when gray-level ‘‘format’’ can be interfaced with an image-processing ‘‘front
is altered (Bruce et al. 1994), but not a¤ected by a end,’’ based on PCA of facial images. Face images are
change in image from color to gray scale or vice represented by the values of a few of their compo-
versa (Bruce, Terry, and Smith 1998). Repetition nents (eigenfaces), represented as input units and
priming can be shown from faces viewed entirely linked to the face recognition units described above.
incidentally, for example, on subject recruitment Because the Hebbian link update mechanism is used
posters (Bruce, Carson, et al. 1998). throughout this version of the IAC model, repetitions
In their interactive activation and competition of exactly the same image give maximum priming: all
(IAC) model of the stages of person identification, links strengthened during the first presentation are
Burton, Bruce, and Johnston (1990) suggested that subsequently used during the second presentation.
repetition priming arises from the strengthening of Because there is considerable overlap between the
connections between units involved in identifying a principal component (PC) representations of the two
face. Following earlier functional models (and par- images, presentation of two di¤erent images of the
ticularly Bruce and Young 1986), they proposed same person produces some priming; because the
pools of face recognition units (FRUs) corresponding overlap is not total, however, priming is reduced.
to the stored visual representations of faces. These This model seems to provide a natural account of
units are connected via processing links to a pool of the processes involved in repetition priming. Burton,
units called ‘‘person identity nodes’’ (PINs), which Bruce, and Hancock (1999) go on to show that a
code individuals rather than faces and can receive model incorporating both image processing and
information from other recognition domains (e.g., cognitive stages of face recognition o¤ers a much
names or voices). These PINs are themselves con- larger predictive range than models limited to the
nected via a further stage of links to personal in- ‘‘perceptual’’ or ‘‘cognitive’’ domains alone.
formation for each of the persons to be identified.
Burton, Bruce, and Johnston proposed a simple 17.5.2 Burton’s Model of Learning New Faces
Hebbian link update mechanism, one that has the (IACL)
e¤ect of strengthening the links used. In this way, if
a face has been seen recently (and the link between a Although our discussion of repetition priming in
particular FRU and the corresponding PIN therefore section 17.5.1 may seem a digression from the main
strengthened), processing will proceed faster on the focus of this chapter, one of the few computational
second presentation of the same face. (For details of models of face learning available relies on the idea
Learning New Faces 333
that exactly the same processes are involved in new sentations, such space is achieved by having redun-
face learning and in priming. dancy in the number of units available. In the IACL
Burton (1994) proposes that priming observed in model, these unused representational patterns are
behavioural experiments is simply the residue of a connected with very small randomly varying con-
mechanism whose chief purpose is to learn faces in nections to the units coding input dimensions (see
the first place. Of the criteria a model of face learn- section 17.5.3). When an input pattern is known, it
ing should satisfy, Burton emphasizes three: such will traverse strong, well-used links and cause exci-
learning should be modeled as automatic, gradual, tation of units corresponding to a known person.
and cumulative. By ‘‘automatic,’’ Burton means that When, however, the pattern is not known, no par-
learning should be unsupervised: the same proce- ticular such unit will become active. Instead, a ran-
dures that allow us to recognize a known face should dom ‘‘spare’’ unit will become selected as the most
also allow us to recognize a face as unfamiliar, and to highly active unit. Simple Hebbian update on all
begin to learn it (cf. chapters 18.4, 20 for types of links in the model will ensure that this new unit will
learning models). The argument is that there is no come to specialize on the new pattern. After several
humunculus in our brains who knows the right presentations, the links connecting the input pattern
answer and can alert us to faces that are or are not and the new person will become as strong as the
genuinely new; instead, the process appears to be links connecting previously known faces.
automatic. Second, the model should allow for The procedure of recruiting new units automati-
gradual learning. We mentioned above that there are cally is, on the surface, cognitively rather odd.
grades of familiarity; intuitively, we appear to know However, it is mathematically exactly the same pro-
some people’s faces better than others. And third, cedure as recruiting a new pattern in a distributed
face learning in the model should be cumulative; we representation. It also has the advantage that it meets
seem to learn new faces with no consequent e¤ects all three criteria set out for it: it is automatic, gradual,
on representations of known faces. The IACL model and cumulative. (For full simulations see Burton
(IAC with Learning) is intended as a model of adult 1994; for a detailed discussion of the computational
face learning, not developmental face learning. This e‰ciency of the procedure, as compared to dis-
final criterion is particularly important in compari- tributed systems, see Burton 1998.)
sons with some connectionist models of learning. We have not yet applied the theoretical model of
Indeed, many connectionist models capture the tran- learning based on hypothetical face patterns to the
sition from no knowledge whatever to total knowl- realistic face images that form the input to the most
edge, for example, by presenting the to-be-learned recent version of the IAC model (Burton, Bruce,
corpus blocked many times (e.g., Farah, O’Reilly, and Hancock 1999). That version takes PCA-coded
and Vecera 1993). The IACL model is an attempt to faces as its input, based on individual static views of
capture a more realistic process in which new faces faces. A further challenge will be to consider how
are learned throughout adulthood. image sequences with their variation over time can
In the IACL model, the pools of FRUs and PINs be coded and learned in models of this type. At this
are large enough to allow space for many more rep- much later stage, we will be able to see whether this
resentations than are used at any one time, whereas, kind of model reproduces the same e¤ects of famil-
in models using both localist and distributed repre- iarization we find in human learning of new faces.
334 Vicki Bruce and Mike Burton
For now, we present one additional thought about example, we know that this continues to be Fred
the kind of learning algorithm that may be appro- because the remainder of Fred remains in the same
priate to model learning new faces. place even as his face pattern varies before us. Fred
has a body and Fred has a voice.
17.5.3 Need Learning Be Unsupervised? 3. When a face is presented to us, we rarely see it
without some accompanying information about who
Burton’s learning algorithm (1994) involves the the person is (their character in a film; the fact that
gradual association of conjunctions of face features they live in your street when met at home, and so
with face recognition units by the simple, unsu- forth). In the terms of the IACL model, activity will
pervised mechanism of Hebbian update, which is be being stimulated and changed in the semantic in-
comparable to the gradual emergence of a represen- formation units and their connections back via the
tation for a familiar face from repeated encounters PINs to the developing face recognition units. Thus
with its visual image. In the simulations of IACL to the process of acquiring new face representations
date, the input ‘‘feature’’ units for learned and new will have top-down support from a range of other
faces remain constant in repeated encounters. An sources of information. In our own future program
obvious need is to extend IACL to a situation that of research, we will investigate whether such non-
convincingly models the process of learning repre- visual constraints a¤ect the way that visual repre-
sentations across variations of the input pattern as sentations are learned.
viewpoints, expressions, and lightings change. Given
our observations of unfamiliar face matching— We have argued that image-based representations
where frequent confusions are made between di¤er- of each new face encountered become consolidated
ent individuals when the task is to match one view through a process of abstraction. This process can be
of an unfamiliar face with another—how can such a implemented within our IACL framework, though
learning system possibly cope with such variation? we have yet to extend this to varying exemplars and
Importantly, the system described by Burton is doing have yet to explore the IACL mechanism using real
a harder job than it strictly needs to. When we en- images of faces, as in Burton et al. 1999. We expect
counter new faces in the real world, there is plenty that familiarization with new faces, once imple-
of scope for supervised learning. To mention just mented, will lead to the emerging properties of
three important factors: generalization, categorical e¤ects, and shifts from
external to internal feature processing. However, the
1. As a face changes in its appearance—through representation of dynamic properties from moving
viewpoint and expression—the temporal sequence faces will require a di¤erent or additional means of
provides constraints so that each image viewpoint image coding from that implemented to date.
is only a slight modification on the one before. It
is extremely important that current computational
work on coding temporal sequences be brought to Acknowledgments
bear on psychological models of learning new faces,
and our own future work will do this. Our research is currently supported by grants from
the U.K. Economic and Social Research Council,
2. As a face is presented in all its variations, other
and the Engineering and Physical Sciences Research
aspects of the learning context remain constant. For
Council.
Modeling
IV
This page intentionally left blank
Models of Perceptual Learning
Shimon Edelman and Nathan Intrator
18
Abstract the field, and the models became, if anything, more
complex than those of 1957.
This chapter addresses learning on a general computational Encouragingly, the models are now also more suc-
level, reviewing a series of broadly relevant theoretical cessful in explaining behavior (rather than merely
notions to identify the dimensions along which varieties of
learning can be classified. In particular, we discuss (1) the
predicting the probability of a certain response to a
goals of learning, (2) the mechanisms that can support given stimulus), while giving no undue troubles to
learning, (3) the cues that a learning system can rely on to the psychologists (for an interesting historical per-
improve its performance, and (4) the paradigms or meta- spective on these issues, see Hintzman 1994). Insofar
phors used to describe learning computationally. Our hope as there is progress, it seems to stem mainly from (1)
is that this review of the computational underpinnings of
the improvement in the experimental techniques that
learning will make the relationships among existing models,
including those mentioned elsewhere in the book, more subserve data collection in behavioral and physio-
readily apparent. logical psychology, and (2) the revision of the theo-
retical basis on which models are built. The ‘‘rodent
on a runway’’ example mentioned above serves well
to illustrate both these points. On the theoretical or
18.1 Introduction
conceptual side, the current explanation takes the
route presaged by Tolman (1948) and based on the
A generation ago, mathematical psychology, then the
concept of cognitive maps (O’Keefe and Nadel
premier discipline in charge of modeling behavior,
1978). On the experimental side, the existence of
appeared to be in poor shape. One prominent math-
cognitive maps in the rat brain could not have been
ematical psychologist, William Estes (1957, 609),
demonstrated without modern multiple-electrode
described the situation: ‘‘Look at our present theories
recording methods and the information-processing
. . . or at the probabilistic models that are multiply-
tools that accompany them.1
ing like overexcited paramecia. Although already too
In this chapter, we shall concentrate on ap-
complicated for the average psychologist to handle,
proaches to the modeling of perceptual learning,
these theories are not yet adequate to account for the
rather than on its phenomenology or on specific
behavior of a rodent on a runway.’’ During the fol-
models—the standard fare of the reviews one finds
lowing decades, when mainstream psychology un-
in the literature (Gibson 1969; LaBerge 1976; Walk
derwent a major paradigm shift, the modeling of
1978; Barto 1989; Gallistel 1990; Gluck and Granger
perceptual learning fared better than what one might
1993; Berry 1994; Gilbert 1994; Sagi and Tanne
have expected from the view expressed by Estes. A
1994; Ahissar and Hochstein 1998). In perceptual
new theoretical outlook, which encouraged thinking
learning, of course, periodic reviews are as important
now termed representational or computational, took over
as in any other discipline blessed with a steady stream
338 Shimon Edelman and Nathan Intrator
of empirical findings. Such reviews stress the relative nisms is between merely exercising memory, on the
merits of learning mechanisms, often at the expense one hand, and using experience with familiar stimuli
of computational theory itself (Marr and Poggio or problems to process or solve new ones, on the
1977). This preoccupation with mechanisms reflects other.
the classical methodological stance, codified by
Popper (1992), according to which empirical studies 18.2.1 Memorization
should begin with a discussion of the models to be
tested and should end by refuting some of the models. Early quantitative studies of the acquisition of de-
It is indeed easier to refute a specific model, clarative information used lists of items to be mem-
mechanism, or wiring diagram than to gain support orized as stimuli. For example, the subjects could be
for a general theory. Nevertheless, a field of study asked to memorize lists of nonsense syllables, whose
stands to gain more from the latter endeavor: by recall was subsequently tested by the experimenter.
providing an explanation for the observed phenom- In this setting, popularized by Ebbinghaus (1885),
ena, a good theory can subsume an entire range of rehearsal of the stimulus is certainly a sensible strat-
models of the underlying mechanisms within the egy, repetition being the mother of learning (see
same formal framework (Deutsch 1997). Follow- chapter 17 for repetition learning of faces). Likewise,
ing this line of reasoning, to understand perceptual in learning that can be described as procedural (i.e.,
learning, we must, first and foremost, address basic learning to perform a perceptual discrimination or a
questions such as ‘‘What does it mean, from a general motor task), repetition was found early on to bring
information-processing standpoint, for a system to about an improvement of performance (see discus-
learn something?’’ Note that the answer ‘‘To learn a sion of Volkman’s 1858 study of cutaneous spatial
perceptual task means to acquire an adequate low- acuity in Gibson 1969).
dimensional internal representation of the stimulus In the century and more since the pioneering
set’’ would be proper in this case (even if it event- work of Volkman and Ebbinghaus, repetition was
ually proved to be factually wrong) because it is shown to lead to improved performance in virtually
coached in general information-processing compu- every perceptual and motor domain tested; at the
tational terms (see Marr and Poggio 1977). By com- same time, the scope of memorization as a paradigm
parison, the answer ‘‘Learning something means for learning was shown to be limited. Specifically, it
growing extra dendrites’’ would constitute a category became clear that performance gain from repetition
mistake—as would the seemingly more abstract an- is transferred only partially to novel situations (Gibson
swer ‘‘Learning is the recruitment of extra memory,’’ 1941; Ellis 1965). The extent and the nature of the
unless the need for this memory is explained in func- transfer depends on the relationship between the sets
tional terms that are algorithm and implementation of perceptual stimuli (or the repertoire of move-
neutral. ments in motor learning) and between the tasks
defined over these stimuli in the original and novel
situations (Osgood 1949). To cite some relatively
18.2 Goals of Learning recent examples, limited transfer was reported by
Fiorentini and Berardi (1981), who found that prac-
The main distinction at the task level that concep- ticing discrimination between spatial contrast gratings
tually precedes any discussion of learning mecha- at one orientation does not improve the performance
Models of Perceptual Learning 339
at an orthogonal orientation (see Fiorentini and Ber- recall, inference ( prediction) of unobserved proper-
ardi, chapter 9, this volume). An analogous situation ties, and valid generalization from one stimulus to
prevails in motor learning; for example, an acquired another (Quine 1969; Dretske 1995).
ability to perform precise elbow flexions was found In psychology, empirical data gathered in the
to transfer only partially from one set of joint angles 1940s and 1950s prompted Guttman (1963) to view
to another (Gottlieb et al. 1988). stimulus generalization as a central theoretical chal-
The distinction between memorization and trans- lenge. Shepard (1987) responded to the challenge by
fer is of crucial importance to any theory of learning. proposing a ‘‘universal law’’ to describe the quanti-
A theory that fails to make this distinction succumbs tative relationship between the likelihood of two
to the same confusion that surrounds the much- stimuli receiving the same response and their per-
publicized inability of an early neural network mod- ceived similarity. More precisely, Shepard showed,
el of learning, the perceptron (Minsky and Papert on the basis of data from a wide range of perceptual
1969), to solve the ‘‘exclusive OR’’ (XOR) problem experiments, that stimuli in each experiment could
(figure 18.1). This problem is special: every nearest be arranged in a low-dimensional metric feature space
neighbor of each input belongs to the opposite class, so that the probability of generalization between any
thus no cluster structure (i.e., no similarity structure, two stimuli was monotonic in their proximity (i.e.,
where nearby points belong to the same class) exists. similarity). Shepard’s treatment of this issue included
The prospects of perceptrons (and of any models a derivation of the monotonic dependence law from
of perceptual learning that share their limitations) some basic assumptions on the probability measure
would be indeed bleak if real-life scenarios tended used to quantify generalization.
to resemble the XOR setup, in which generalization In theoretical neurobiology, generalization under-
is ill defined (Bishop 1995). As we shall argue next, lies the ‘‘fundamental hypothesis’’ of Marr’s theory
however, learning scenarios that focus on memori- of the cerebral neocortex (1970, 150–151): ‘‘Where
zation—testing subjects (1) in a fixed task and (2) instances of a particular collection of intrinsic prop-
with the same stimuli encountered during the learn- erties (i.e., properties already diagnosed from sensory
ing phase—cover only a part of the great variety of information) tend to be grouped such that if some are
everyday situations in which learning is known to present, most are, then other useful properties are
occur. likely to exist which generalize over such instances.
Further, properties often are grouped in this way.’’
18.2.2 Generalization Although this hypothesis seems at present every bit
as convincing as it must have appeared to Marr, it
The behavioral importance of transfer of learning, remains, unfortunately, empirically unsubstantiated;
or generalization, to novel conditions (as contrasted its vindication or refutation is likely to bear on sta-
with memorization) has been pointed out and dis- tistical theories of brain function, such as those of
cussed by philosophers, psychologists, and neuro- Uttley (1959) and Marr (1970), and the neural net-
biologists. In philosophy, the ‘‘naturalistic’’ approach work theories of their recent successors (more on
to epistemology (Kornblith 1985) involves the con- this below).
cept of natural kinds—categories of objects that The foundational status of generalization in visual
share su‰ciently many features to support associative perception and cognition can be easily illustrated
340 Shimon Edelman and Nathan Intrator
Figure 18.1
(Upper left) Truth table definition of a two-variable exclusive OR (XOR) problem. (Lower left) Note how the points
belonging to the two classes are interspersed among each other. The common characteristic of this kind of problem, which
impedes generalization, is that the probability of two neighboring points belonging to the same class is at chance. (Upper
right) Outlines of size and position of two Gaussian receptive fields (RFs; see Poggio, Edelman, and Fahle 1992) in vernier
discrimination problem. (Lower right) Representations of 200 vernier stimuli in the space of the outputs of these same two
RFs. The simulation that produced this plot used a 100 100 image, and two Gaussian filters with s ¼ 30, positioned at
ð70; 30Þ and ð30; 70Þ. The two symbols, and , correspond to the two senses of vernier displacement, making this a class-
conditional probability density plot of a sort. The vernier displacement in this experiment ranged from 5 to 15 pixels. The
crucial characteristic here are the clusters of points belonging to the same class, which are simply connected and unimodal
(human observers find learning more di‰cult when the class-conditional distributions are disjunctive or multimodal;
Flannagan et al., 1986). It seems safe to conjecture that in general the class-conditional densities arising from perceptual tasks
can be relatively easily made to look like this, facilitating decision making (unlike the case of the XOR example). In this
chapter, we argue that learning can be construed as the formation of a representation space in which the problem at hand is
well-behaved in this sense.
Models of Perceptual Learning 341
in intuitive terms, on an everyday task. Consider, for 1983), the components of neural models of learning
example, learning to recognize a face from several are activation states of simple computing elements
snapshots. The observer’s ability to recognize the face and their interconnection patterns (Selfridge 1959;
in this case most probably extends to new images Hinton 1989; Rumelhart and Todd 1993).
(obtained, say, under various combinations of view- It is now widely acknowledged that the principles
point and illumination). Moreover, the observer is of operation of neural learning models apply also to
also expected to be able to solve a range of percep- more traditional computational paradigms and data
tual problems involving that face (e.g., to estimate its structures (Omohundro 1987). Even more impor-
direction of gaze, to categorize its various expres- tantly, neural networks turn out to be amenable to
sions, etc.). These latter abilities e¤ectively require mathematical analysis that invokes well-established
that learning be transferred from one set of stimuli statistical tools dealing with inference and decision
(i.e., images of the many faces previously processed making (Bishop 1995). For example, Widrow’s
by the subject) to another (i.e., images of the new ‘‘Adaline’’ (adaptive linear element) networks (1985)
face). Thus any theory of perceptual learning must can be identified with linear discriminant functions,
include a component that would account for gener- and multilayer perceptrons with multivariate multiple
alization across stimuli and across tasks, over and nonlinear regression. (For further parallels between
above rote memory (figure 18.2). The central role of the neural network terminology and that of statistics,
generalization in learning underscores the impor- see Sarle 1994.)
tance of experiments such as those of Fiorentini and Inferential statistics thus constitutes a useful foun-
Berardi (1981), which define the limits of general- dation for the understanding of the computational
ization in the human perceptual system, and thereby capabilities of neural networks. If this foundation is
make a crucial contribution to the discovery of the to be useful in the development of specific models of
principles and the mechanisms that support it. learning in the nervous system, statistical samples of
stimuli must be shown to contain information nec-
essary for learning. Having been downplayed for
18.3 Mechanisms of Learning decades by Chomsky and his school, the notion that
statistical inference can support learning even in
The characteristics of learning discussed thus far have markedly ‘‘symbolic’’ domains such as language ac-
to do with the nature of the task, where the main quisition is now making a comeback. This process is
distinction is between memorization and generaliza- aided by the growing evidence that humans ( both
tion. We now consider the mechanisms employed adults and infants) are sensitive to statistical cues pre-
by models of learning to explain the improvement of sent in linguistic stimuli. For example, from such
the performance with practice. At the highest level cues, subjects can implicitly extract information
of abstraction, a common (though not entirely war- about boundaries between the underlying morpho-
ranted) distinction is between symbolic mechanisms logical units (Sa¤ran, Aslin, and Newport 1996),
and neuromorphic ones. Whereas in symbolic word meaning (Markson and Bloom 1997), and
learning the building blocks of models are proposi- even grammarlike rules (Berns, Cohen, and Mintun
tions and rules (Carbonell, Michalski, and Mitchell 1997). Consequently, models built around symbolic
342 Shimon Edelman and Nathan Intrator
Figure 18.2
(Top) Schematic diagrams of three varieties of data- and task-related situations in learning. (A) When inputs and required
outputs are the same in both training and testing phases, learning amounts to memorizing the input-output association, such
as that of a name with a face. The arrow leading from the train to the test box represents the parameters acquired by the
adaptive mechanism during the learning process. (B) If new data are to be mapped into the same output space, the system
must generalize the previously learned association, such as by naming a familiar face seen under novel conditions (e.g., pe-
culiar illumination). (C) If both the input and the output spaces change between training and testing, the problem turns into
that of transfer of learning to a new task, such as matching two views of an unfamiliar face on the basis of prior experience
with other face stimuli. (Bottom) Schematic diagram of the dependence of transfer on the relationship between the charac-
teristics of two tasks. (Adapted from Osgood 1949.) The degree of transfer grows with the similarity between the stimuli in
the two tasks, and, for highly similar stimuli, is reduced if the required responses are di¤erent.
Models of Perceptual Learning 343
Interestingly, the attempts on the part of a learn- measurement space—largely preserves the local ge-
ing system to improve its performance under the ometry of the distal problem space (Edelman 1999).7
limitations imposed by its architecture are precisely The first reason can be illustrated by a series of
what may pass as the manifestation of the process examples taken from all areas of perception. Con-
of learning to an external observer. For example, a sider, for example, the vernier discrimination task in
system constrained to use RBFs may start by ‘‘tiling’’ which the observer is to judge the sense of the rela-
the inside and the outside of the circle in figure 18.4, tive displacement of two abutting line segments (see
using up many basis function units. It may then shift figure 18.1, right). The solution in this case is para-
to the more economical representation in which the meterized by a single variable, which controls the
inside and thus the outside of the circle are repre- displacement of the segments perpendicular to their
sented by a single basis—an event that would look extent. Stepping this variable by small increments
like the kind of feature discovery mentioned by through a range of values between, say, 15 00 and
Gibson. In other words, learning can be defined þ15 00 would cause the measurement space represen-
computationally as the art of creating the most suit- tation of the resulting stimulus to ascribe a one-
able representation of the data, given the constraints dimensional manifold.8 All that a visual system
of the model at hand.6 would have to do to learn vernier discrimination
would be to interpolate this manifold from a set of
18.5.2 Learning Visual Manifold Geometry: examples (i.e., input-output pairs), as described in
Regression section 18.4.2.
As another illustration, we may consider the prob-
For a resource-constrained system, attempting to lem of learning to recognize an object from examples
solve the learning problem in the original feature (i.e., a few of its stored views, each view being con-
space may prove too complicated, as illustrated by strued as a snapshot of the multidimensional mea-
the application of RBFs to the circle example. At the surement space). For a rigid object allowed two
same time, attempting to remap the original problem rotational degrees of freedom, corresponding to the
into a new representation space where it would two axes of rotation in depth, the manifold spanned
become linear may be equally hard, rendering the in the measurement space will be two-dimensional,
problem intractable. Fortunately, it appears that many and will be amenable to learning from examples, as
perceptual problems possess an inherent structure that shown in Poggio and Edelman 1990. An entire range
makes them amenable to learning methods that rely of other learning tasks having to do with object rec-
neither on the exhaustive tiling of a high-dimensional ognition can be solved on the basis of related princi-
space (intractable because of the curse of dimension- ples, as shown in Edelman 1999 and Duvdevani-Bar
ality) nor on a sophisticated remapping ( possibly et al. 1998.
beyond the visual system’s capabilities). The struc- In psychology, the well-behavedness of the inter-
ture in question is that of a smooth low-dimensional nal representation space of various visual qualities has
manifold, which arises for two reasons: (1) problem been noted repeatedly, beginning with the shift to-
spaces in a typical perceptual task are parameterized ward representational theories of vision in the 1960s.
by only a few variables (Edelman and Intrator 1997), The concept of an internally represented stimulus
and (2) the ‘‘front end’’ of a typical visual system—its space was mentioned by Guttman (1963, 144), who
348 Shimon Edelman and Nathan Intrator
pointed out in a discussion of generalization in animal the dimensionality of the embedding space. This in-
learning that the pigeon ‘‘knows the spectrum, in an cludes the self-organizing map algorithm (Kohonen
important sense of the word ‘know’ ’’—it exhibits 1982), and the di¤erent varieties of autoencoders or
the kind of orderly generalization between colors that bottleneck networks (Cottrell, Munro, and Zipser
psychologists routinely observe, and must, therefore, 1987; Leen and Kambhatla 1994).
possess an internal ‘‘color space.’’ More than two The performance of such unsupervised or self-
decades later, Shepard (1987) formulated his law of supervised manifold-extracting algorithms can be
generalization in terms of proximities in an internal improved if additional knowledge is brought to bear
psychological space, observing that structure inherent on the problem. Typically, this is done by making
in various distal ‘‘quality spaces’’ (Clark 1993), such the learning mechanism observe certain invariances
as the color continuum discussed by Guttman, is known to apply to the problem (Földiák 1991; Wis-
faithfully represented internally. kott 1998). A particularly simple way to do that is
It is interesting to compare the observations con- to provide the label of the category to which each
cerning the low dimensionality and the smoothness stimulus belongs.9 To see how this information helps
of the internally represented quality spaces to similar the algorithm isolate the relevant manifold, note that
observations made by statisticians and neural net- directions orthogonal to it can be e¤ectively speci-
work researchers. In nonparametric statistics, for fied by forcing stimuli that di¤er along those direc-
example, the surprisingly good performance of tions to be mapped to the same category (Intrator
nearest-neighbor methods, which rely on raw fea- and Edelman 1997).
ture space distances, has been explained intuitively:
the relevant points in these spaces—that is, the 18.5.3 Learning Visual Category Structure:
examples—tend to be confined to smooth low- Classification
dimensional subspaces (Friedman 1994). In neural
network research, an analogous observation can be From the perspective of the task, the main di¤erence
found in Bregler and Omohundro 1995. between regression and classification is that in
Realizing that the problems at hand typically regression the location of the point within the low-
possess such a convenient structure is, however, only dimensional structure matters, whereas in classifi-
the first step toward their solution. Lowering the cation it does not. For example, the location of the
dimensionality of the space in which the stimuli are point representing a face in a face space (the mani-
originally encoded is a nontrivial operation; as we fold corresponding to the di¤erent possible views of
pointed out above, in human visual perception the the same face) would encode its orientation, which is
original dimensionality of any stimulus is, nominally, a piece of information that should not be discarded.
on the order of 10 6 , which is the number of fibers in In comparison, in the vernier task, where the prob-
each optic nerve. The illustration (figure 18.3, left) lem is that of classification, only the membership in
of the manifold embedded in two-dimensional one of the two clusters in the representation space
space, is thus highly simplified. Many of the com- matters to the system.
putational approaches devised for the extraction of Despite this di¤erence, the basic considerations
low-dimensional manifolds do not scale well with identified before in the discussion of regression apply
Models of Perceptual Learning 349
also to classification. In particular, the curse of consists of the centroids of the two clusters of data
dimensionality still has to be taken into account. points).
Huber (1985) illustrates this point quantitatively, Thus both regression and classification can be
by showing how di‰cult it is to find a three- subsumed under a common framework, which calls
dimensional Gaussian bump (which could, in terms for estimating the joint probability density of all the
of figure 18.3, right, correspond to one of the class- variables included in the data set. It is well known
conditional clusters), when it is embedded in a that this information reveals everything there is to
ten-dimensional space. Although neurally inspired know about stochastic data, such as the measure-
models of learning that are tailored specifically for ments performed by a perceptual system on the
categorization and mixture estimation do exist world. Although the underlying generator of the
(Carpenter and Grossberg 1990; Carpenter et al. data (thus the quantities needed for regression or
1992; Williamson 1997), they are not expected to classification) can then be estimated optimally from
deal better with realistically high-dimensional cases the density function, the first step in this process—
than the knowledge-based models mentioned ear- the inference of an unconstrained density function
lier (which were designed for manifold extraction, from data—is prone to the curse of dimensionality,
yet should be equally capable of clustering). as shown in the seminal work of Stone (1980, 1982).
In view of this problem, researchers typically take
18.5.4 Learning Joint Input-Output Probability two approaches, which are not mutually exclusive.
Density The first is to make some assumptions about the
density function. For example, they may assume that
When learning is treated as a problem in statistical the density function is smooth, then estimate it using
inference, the observations we made in section 18.5.3 splines (Wahba 1979) or radial basis functions (Poggio
can be rephrased using the concept of the underlying and Girosi 1990). They may instead assume some-
generator of the data—the entity that causes what- thing about the structure of the density. For example,
ever regularities are present in the data set. In visual it may be postulated to belong to an additive model,
perception, this entity is the distal stimulus, which making it expressible as a sum of functions of some
gives rise to the observed values of features through low-dimensional projections of the data (Stone 1985,
a complex process (reflection and scatter of light, 1986). Or they may assume that the density is facto-
propagation in the medium, refraction by the optics rial, namely, a product of marginal densities of one
of the eye, phototransduction, etc.). In regression- variable (Dayan, Hinton, and Neal 1995). The latter
like tasks, the distal stimulus space is continuous by two methods, though they do not lower the dimen-
nature, for example, the continuum of views of an sionality of the density function, make the estimation
object that undergoes rotation in front of the ob- process more e‰cient and less prone to the curse of
server (in figure 18.3, left, it is the smooth curve dimensionality.
underlying the sausagelike cloud of points). In clas- The second general approach, which bypasses the
sification, the distal stimulus space is discrete, for problem of density estimation, is based on the ob-
example, the set of categories to which the viewed servation that, for many practical problems, only a
object may belong (in figure 18.3, right, this space certain function of the density is required. The hope
350 Shimon Edelman and Nathan Intrator
is that such a function can be easily computed di- 18.6.1 On the Levels of Explaining Learning
rectly from the data, without having to make the
full density estimation, which happens, for example, The overarching concern in the modeling of a per-
when the desired function is defined over a low- ceptual phenomenon is getting the performance
dimensional manifold embedded in the original space right. Beyond that, however, there is a considerable
or, more generally, when the desired function has a variation in what is deemed acceptable: whereas some
simpler structure compared to the full density. In such comprehensive models treat both the computational
cases, the learning system may attempt to extract the (theoretical) and the implementational aspects of the
low-dimensional representation of the problem from problem, others tend to concentrate on the issues
the data, using an unsupervised approach such as of implementation and mechanism. Models built
principal component analysis and its generalizations, around neural networks most likely belong to the
or using a supervised approach tailored to the desired second group, going straight from the phenomenol-
target function, as in many feedforward network ogy to a hypothesis about the underlying mecha-
models. nism, perhaps also attempting to emulate along the
In all these cases, a model would do well by way the real biological neural network.
applying the methods listed in sections 18.5.2 and We illustrate this observation with a striking ex-
18.5.3, which dealt with learning manifold extrac- ample of perceptual learning, found in the task of
tion (regression) and clustering (classification). On detecting a small low-contrast Gabor patch projected
the other hand, those methods cannot be practically onto a certain retinotopically defined location. The
subsumed under the aegis of density estimation un- detection threshold in this task depends on whether
less the estimation algorithm (1) aims for learning a the target patch is flanked at a distance by patches
certain target function of the density, which is usu- of similar orientation and spatial frequency (Polat
ally problem specific, and (2) relies on some prior and Sagi 1993). The e¤ect of the flanking patches is
assumptions about the properties of the desired rep- amenable to learning: the spatial range of the e¤ect
resentation, such as low dimensionality and smooth- (i.e., the maximum e¤ective distance between the
ness (Intrator 1993). target and the flanking patches) grows with practice
(see Zenger and Sagi, chapter 10, this volume). Sig-
nificantly, learning is only possible if the original,
18.6 Discussion
untrained range is extended gradually by exposing
the subject to configurations of progressively larger
The theoretical stance adopted thus far equates
and larger extent (Polat and Sagi 1994).
learning with the acquisition of e‰cient representa- A phenomenon such as this seems to positively
tions, a computational procedure that can be regarded
demand an explanation at the mechanism level, in-
as a kind of statistical inference. Although we may
voking receptive fields of retinotopic ‘‘units,’’ linked
seem to have strayed far from the gritty details that
laterally and exerting facilitatory influence on each
must be dealt with by any model aiming to simulate
other. Polat and Sagi (1994) o¤ered just such an ex-
human learning behavior, we believe that a good
planation for their psychophysical findings. As we
model starts at the top, with a clear notion of what is
claimed in section 18.1, however, because they
being modeled and why.
concentrate on the wiring details at the expense of
Models of Perceptual Learning 351
leaving the computational goal of the system out of tion may benefit from exploring the possibilities
the picture, models formulated primarily in the lan- originally developed in the context of object recog-
guage of units and connections achieve less than nition (e.g., interpolation with feedforward basis
what a model can and should achieve. To support functions).
this argument, let us reconsider the ‘‘lateral learning’’ An edifying perspective on the issue of levels of
scenario, keeping in mind the taxonomy of learning modeling is provided by recalling some of the ‘‘old-
paradigms discussed earlier. fashioned’’ models of brain function (and learning)
Assume for the moment that the goal of the sys- produced by neurobiologists. Two such models that
tem is to detect the faintest possible line element were prominent in their own time, one dealing with
(a real-life counterpart to a Gabor patch) in a given ‘‘universals’’ or the problem of invariance (Pitts and
retinal location. Merely lowering the decision McCulloch 1947/1965) and the other with proba-
threshold for that location will likely just increase the bilistic generalization (Marr 1970), actually did link
false-alarm rate there; additional information must theory and mechanism. Marr’s model of the neo-
be brought to bear on the decision if it is to be reli- cortex (1970), for example, spans the entire possible
able. The presence of other line elements in the vi- range of levels. It starts with a general, yet succinctly
cinity would count as the necessary additional support phrased hypothesis concerning the probabilistic
if they are compatible with the original hypothesis structure of the world (the ‘‘fundamental hypothesis,’’
(i.e., if their orientation is consistent with that of the mentioned in section 18.2.2), and ends with a
element whose fate they are about to seal). Thus the detailed explanation of the possible ways in which
task at hand can be reformulated as that of (literal) neuroanatomy and neurophysiology of the cortex
interpolation between the flanking lines, or of ex- may be tuned to put the observed probabilities to
trapolation if the continuation of an ‘‘end-stopped’’ work. This style of modeling, which integrates dif-
segment is sought. ferent levels of explanation, requires a combination
This formulation makes it possible to uniformly of encyclopedic knowledge with considerable inge-
treat a range of perceptual learning tasks. Indeed, on nuity on the part of the modeler. Unfortunately, it is
an abstract level, learning to detect a Gabor patch now quite rare, having been replaced by a method-
flanked by similar patterns is now seen to be the same ology that allows the levels—computational, algo-
as learning to recognize an object ‘‘sandwiched’’ be- rithmic, and implementational—to be kept separate.
tween two familiar views from a novel viewpoint.
The analogy drawn between these two tasks hinges 18.6.2 Prognosis
on a parallel between the view space of the ob-
ject, on the one hand, and the ‘‘space’’ space—that In visual perception, learning is a pervasive phe-
is, the retinal location space—of the Gabor patch, nomenon, which, when properly studied, o¤ers the
on the other. Once this analogy is accepted, cross- researcher a unique searchlight on the inner work-
fertilization may occur in both directions. Whereas ings of the system. Although it would be rash to
models of object recognition may benefit from pos- predict what this searchlight might reveal, there
tulating a mechanism that carries out interpolation are four strategics we would like to see adopted in
by growing lateral links between neighboring units modeling perceptual learning:
in a view representation space, models of line detec-
352 Shimon Edelman and Nathan Intrator
1. Integrate past achievements. Attempts to develop both on the level of behavior, and on the level of
mathematical models of learning date back more than architecture.
half a century. Much of the work carried out before 4. Go after the big question. To understand the brain,
mid-1960s has now been branded ‘‘behaviorist’’ and we need a really comprehensive explanation, one
e¤ectively buried in the libraries. Reexamining that that starts from a concrete premise, yet spans all the
work may lead to interesting insights into the nature levels of the ‘‘hierarchy’’ of computation, represen-
of the present-day models (see Hintzman 1994). tation, algorithm, and implementation—a postulate
2. Look at learning di¤erently. A diametrically oppo- about what the brain actually does. Several such
site trend is that of complete rejection of both the postulates are available, for example, Marr’s proba-
old and the contemporary models of learning in bilistic inference (1970), Barlow’s redundancy re-
favor of esoteric theories that involve concepts such duction (1990), and Poggio’s function approximation
as catastrophes, self-organized criticality, or phase (1990). A more intense competition in this arena is
transitions in dynamical systems. Although the trend likely to lead to some exciting developments in the
is in response to legitimate challenges, such as the modeling of learning.
need to explain abrupt learning and related phe-
The ability to learn, at all levels and under all cir-
nomena (see Rubin, Nakayama, and Shapley, chap-
cumstances, is the most striking attribute of human
ter 13, this volume), we believe that perceptual
cognition. What would it take to really understand
learning research will be best served by the widest
it? Just as Gounod’s Faust asked Mephistopheles for
variety of di¤erent approaches, from the very tradi-
youth—the treasure that contains all others—so we
tional to the very novel.
wish for a model of the brain that would make the
3. Explain as much as possible. The ‘‘dynamical’’ modeling of perceptual learning superfluous.
models attempt to explain the behavior of the per-
ceptual system by appealing to an isomorphism be-
tween its physics (i.e., the di¤erential equations that Notes
describe it) and the physics of other systems exhibit-
ing a similar behavior. In that, they resemble the 1. Cognitive maps in the rat brain are thought to reside
in the hippocampus, a cortical structure implicated in per-
behaviorist models, which skirt the issues of repre-
ceptual (spatial) and other kinds of learning. Information
sentation, and deal with disembodied equations about the spatial location of the animal turns out to be
aimed at mimicking the phenomenology of the tar- represented in the firing patterns of hippocampal ‘‘place’’
get system. Nor are they alone in this regard. Purely cells, whose ensemble activity constitutes an internal
representational models also end up dealing only cognitive map of the rat’s environment (Wilson and
with the phenomenology; a good example is She- McNaughton 1993). Another class of cells in the hippo-
campus are the ‘‘head direction’’ cells, which serve as an
pard’s law of generalization (1987; mentioned in
internal compass to orient the cognitive map. The func-
section 18.2.2), which makes no claims as to the re- tional properties of place and head direction cells emerge
ality of the ‘‘psychological similarity space’’ it postu- from a complex and as yet poorly understood interaction
lates. In contrast to all these, explanations o¤ered by between internally generated, self-motion cues (e.g., ves-
the more daring connectionist models (which bite tibular information) and external sensory input (e.g., visual
landmarks; Knierim, Kudrimoti, and McNaughton 1995).
the bullet and hope for the best) include parallels
Models of Perceptual Learning 353
a global spatial frequency analysis (Hancock, Badde- images. Imagine that each image patch, represented
ley and Smith 1992). Thus, second-order statistics by the vector x, has been formed by the linear
alone do not su‰ce to predict the formation of combination of N basis functions. The basis func-
localized edge detectors. tions form the columns of a fixed matrix, A. The
Additional constraints are required. Field (1987, weighting of this linear combination (which varies
1994) has argued for the importance of sparse, or with each image) is given by a vector, s. Each com-
‘‘minimum entropy,’’ coding (Barlow 1994), in ponent of this vector has its own associated basis
which each feature detector is activated as rarely as function, and represents an underlying ‘‘cause’’ of the
possible. This has led to feature-learning algorithms image. The linear image synthesis model is therefore
with a ‘‘projection pursuit’’ flavor (Huber 1985, given by:
Intrator 1992, Baddeley 1996, Olshausen and Field
x ¼ As ð19:1Þ
1997).
An alternative constraint is to start with an infor- which is the matrix version of the set of equations
mation-theoretic criterion that maximizes the joint
entropy of a nonlinearly transformed output feature X
N
xi ¼ aij sj ð19:2Þ
vector. This is the approach taken by ‘‘independent j¼1
components analysis’’ (Comon 1994) which can
achieve the blind separation of mixed sources ( Jutten where each xi represents a pixel in an image, and
and Hérault 1991; Bell and Sejnowski 1995a, 1996). contains contributions from each one of a set of N
Finding independent components is equivalent to image ‘‘sources,’’ sj , linearly weighted by a coe‰-
Barlow’s redundancy reduction problem; therefore if cient, aij .
Barlow’s reasoning is correct, the independent com- The goal of a perceptual system, in this simplified
ponents should produce filters which are localized framework, is to linearly transform the images, x,
and oriented, and in fact it does. In addition, when with a matrix of filters, W, so that the resulting
applied to natural images, the outputs of the result- vector:
ing filters are more sparsely distributed than those of
u ¼ Wx ð19:3Þ
other decorrelating filters, thus supporting some of
the arguments of Field (1994) and helping to explain recovers the underlying causes, s, possibly in a dif-
the results of Olshausen and Field (1997) from an ferent order, and rescaled. Representing, by P, an
information-theoretic point of view. arbitrary permutation matrix (all zero except for a
We will return to the issues of sparseness, noise single ‘‘one’’ in each row and each column), and, by
and higher-order statistics. First, we describe more S, an arbitrary scaling matrix (nonzero entries only
concretely the filter-learning problem. on the diagonal), such a system has converged when:
Then the basis functions (columns of A) and the the spatial arrangement of the image and flattening
filters that recover the causes (rows of W ) have the its frequency (amplitude) spectrum (Goodall 1960;
simple relation: W ¼ A1 . Atick and Redlich 1993). Example ZCA filters and
All that remains in defining an algorithm to learn basis functions are shown in figure 19.1b.
W (and thus also A) is to decide what constitutes a Another way to constrain the solution is to at-
‘‘cause.’’ We concentrate here on algorithms pro- tempt to produce outputs that are not just decorre-
ducing causes that are decorrelated, and those at- lated but statistically independent ( Jutten and Hérault
tempting to produce causes that are statistically 1991; Comon 1994). The values of the ui are in-
independent. dependent when their probability distribution, f u ,
Q
factorizes: fu ðuÞ ¼ i fui ðui Þ. There are many ICA
algorithms, based on di¤erent approaches (Cardoso
19.2 Decorrelation and Independence and Laheld 1996; Karhunen et al. 1996; Amari,
Cichoki, and Yang 1996; Cichocki, Unbehauen, and
The matrix, W, is a decorrelating matrix when the co- Rummert 1994; Pham, Garrat, and Jutten 1992; Bell
variance matrix of the output vector, u, satisfies: and Sejnowski 1995a).
huuT i ¼ diagonal matrix ð19:5Þ ICA produces decorrelating filters that are sensi-
tive to both phase (locality) and frequency informa-
In general, there will be many W matrices which tion, just as in transforms involving oriented Gabor
decorrelate. For example, when huuT i ¼ I, then: functions (Daugman 1985) or wavelets. These filters
WT W ¼ hxxT i1 ð19:6Þ are thus semilocal, depicted in figure 19.2 as partway
along the path from the local (ZCA) to the global
which clearly leaves freedom in the choice of W. (PCA) solutions in the space of decorrelating so-
There are, however, several special solutions to Eq. lutions. Example ICA filters are shown in figure
(19.6). 19.1d and their corresponding basis functions are
Principal components analysis (PCA) is the shown in figure 19.1e.
orthogonal solution to Eq. (19.5). The principal It is important to recognize two di¤erences be-
components come from the eigenvectors of the tween finding an ICA solution, WI , and other de-
covariance matrix. The filters are orthogonal. When correlation methods: (1) there may be no ICA
the image statistics are stationary (Field 1994), the solution, and (2) a given ICA algorithm may not find
PCA filters are global Fourier filters, ordered accord- the solution even if it exists, because there are ap-
ing to the amplitude spectrum of the image. Exam- proximations involved. In these senses, ICA is dif-
ple PCA filters are shown in figure 19.1a. ferent from PCA and ZCA, and cannot be calculated
If W is forced to be symmetrical, so that WZT ¼ analytically, for example, from second-order statistics
WZ , then the resulting decorrelating filters are (the covariance matrix), except in the Gaussian case
zero-phase (ZCA). ZCA is in several ways the polar (when second-order statistics completely characterize
opposite of PCA. It produces local (center-surround the signal distribution).
type) whitening filters, which are ordered according The approach developed in Bell and Sejnowski
to the phase spectrum of the image. That is, each 1995a was to maximize by stochastic gradient ascent
filter whitens a given pixel in the image, preserving the joint entropy, H½ gðuÞ, of the linear transform
358 Anthony J. Bell and Terrence J. Sejnowski
since for PCA they are the same thing. (b) Zero-phase
components analysis (ZCA, or WZ ): The first six entries
in this column show the one-pixel-wide center-surround
filter which whitens while preserving the phase spectrum.
All are identical, but shifted. The lower six entries
Figure 19.1
ð37; 60 . . . 144Þ show the basis functions instead, which are
Selected decorrelating filters and their basis functions
the columns of the inverse of the WZ matrix. (c) The
extracted from the natural scene data. Each type of decor-
weights, W, learned by the independent component anal-
relating filter yielded 144 12 12 filters, of which we only
ysis network trained on WZ -whitened data, showing (in
display a subset here. Each column contains filters or basis
descending order) the DC filter, localized oriented filters,
functions of a particular type, and each of the rows has a
and localized checkerboard filters. (d ) The corresponding
number relating to which row of the filter or basis function
ICA filters, in the matrix WI , look like whitened versions
matrix is displayed. (a) Principal components analysis
of the W-filters. (e) The corresponding basis functions,
(PCA, or WP ): The 1st, 5th, 7th, etc. principal compo-
columns of WI1 (or A). These are the patterns that opti-
nents, showing increasing spatial frequency. There is no
mally stimulate their corresponding ICA filters, while not
need to show basis functions and filters separately here,
stimulating any other ICA filter, so that WI A ¼ I.
Learning to Find Independent Components 359
Figure 19.3
Matrix of 144 filters obtained by training on natural images whitened by zero-phase components analysis. Each filter is a row
of the matrix W. The independent components analysis basis functions on ZCA-whitened data are visually the same as the
ICA filters. On nonwhitened data, the filters look like high-pass versions of the filters shown here, and the basis functions
look like low-pass versions of them.
Learning to Find Independent Components 361
Figure 19.4
Log distributions of univariate statistics of the outputs of
independent, zero-phase and principal components analy-
sis (ICA, ZCA, and PCA) filters, averaged over all filters of
each type. All three are approximately double-exponential
distributions, but the more kurtotic ICA distribution is
slightly peakier and has a longer tail, showing that it is Figure 19.5
sparser than the others. This distribution (and the two di- Contour plots of log distributions of pairwise statistics of
mensional ones in figure 19.5), although averaged over the outputs of independent, zero-phase, and principal
the outputs of all filters, are extremely similar to the dis- components analysis (ICA, ZCA, and PCA) filters. (a, c, e)
tributions output by individual filters (respectively, pairs of Joint log distributions averaged over all pairs of output
filters). The only exception is the DC filter (top left in filters of each type, and all images. (b, d, f ) Product of
19.3) which has a more Gaussian distribution. marginal (univariate) distributions. The ICA solution best
satisfies the independence criterion that the joint distri-
bution has the same form as the product of the marginal
In summary, the filters found by the infomax ICA distributions.
algorithm with a logistic nonlinearity are localized,
oriented, and produce outputs distributions of very
work along these lines was by Linsker (1988), who
high kurtosis.
first proposed the ‘‘infomax’’ principle that underlies
our own work. Linsker’s approach, and that of Atick
19.3 Comparisons with Other Approaches and Redlich (1990), Bialek, Ruderman, and Zee
(1991), and van Hateren (1992) uses the second-
A substantial literature exists on the self-organization order (covariance matrix) approximation of the re-
of visual receptive fields through factors such as quired information-theoretic quantities, and generally
learning. Many contributions have emphasized the assumes Gaussian signal and Gaussian noise, in which
roles of decorrelation and PCA (Oja 1989; Sanger case the second-order information is complete. The
1989; Miller 1988; Hancock, Baddeley, and Smith explicit noise model and the restriction to second-
1992; Földiák 1990). Often this has been accom- order statistics mark the two di¤erences between
panied by information-theoretic arguments. The first these approaches and our approach to infomax.
362 Anthony J. Bell and Terrence J. Sejnowski
The technical reason for this failure is that second- are as sparsely distributed as possible. The sparseness
order statistics correspond to the amplitude spectrum constraint is imposed by a nonlinear function that
of a signal (because the Fourier transform of the pushes the activity of the components of u toward
autocorrelation function of an image is its power zero.
spectrum, the square of the amplitude spectrum.) Thus the similarity of the results produced by
The remaining information, higher-order statistics, Olshausen and Field’s network and ours may be
corresponds to the phase spectrum. The phase spec- explained by the fact that both produce what are
trum is what we consider to be the informative part perhaps the sparsest possible ui distributions, though
of a signal, since if we remove phase information by di¤erent means. In emphasizing sparseness di-
from an image, it looks like noise, while if we re- rectly, rather than an information theoretic criterion,
move amplitude information (for example, with Olshausen and Field do not force their ‘‘causes’’ to
zero-phase whitening, using a ZCA transform), the have low mutual information, or even to be decor-
image is still recognizable. Edges and what we con- related. Thus their basis function matrices, unlike
sider ‘‘features’’ in images are ‘‘suspicious coinci- ours, are singular, and noninvertible, making it dif-
dences’’ in the phase spectrum: Fourier analysis of ficult for them to say what the filters are that corre-
an edge consists of many sine waves of di¤erent spond to their basis functions. Recently, Lewicki and
frequencies, all aligned in phase where the edge Olshausen (1999), working with overcomplete rep-
occurred. resentations, have overcome these problems.
As in our conclusions about ‘‘noise,’’ we feel that Our approach, on the other hand, emphasizes in-
a more general information-theoretic approach is dependence over sparseness. Examining figures 19.4
required, an approach taking account of statistics of and 19.5, we see that our filter outputs are also very
all orders. Such an approach is sensitive to the phase sparse. This is because infomax with a sigmoid non-
spectra of the images, and thus to their characteristic linearity can be viewed as an ICA algorithm with an
local structure. These conclusions are borne out by assumption that the independent components have
the results of ICA, which demonstrate the emergence super-Gaussian PDFs. It is worth mentioning that an
of local oriented receptive fields, which second-order ICA algorithm without this assumption will find a
statistics alone fail to predict. few sub-Gaussian (low-kurtosis) independent com-
Several other approaches have arisen to deal with ponents, though most will be super-Gaussian (Lee,
the unsatisfactory results of simple Hebbian and anti- Girolami, and Sejnowski 1999).
Hebbian schemes. Field (1987, 1994) emphasized, Sparseness, as captured by the kurtosis, is one pro-
using some of Barlow’s arguments (1989), that the jection index often mentioned in projection pursuit
goal of an image transformation should be to con- methods (Huber 1985), which look in multivariate
vert ‘‘higher-order redundancy’’ into ‘‘first order- data for directions with ‘‘interesting’’ distributions.
redundancy.’’ These arguments led Olshausen and Intrator (1992; see chapter 18), who pioneered the
Field (1997) to attempt to learn receptive fields by application of projection pursuit reasoning to fea-
maximizing sparseness. In terms of our figure 19.6, ture extraction problems, used an index emphasizing
they attempted to find receptive fields (which they multimodal projections, and connected it with the
identified with basis functions—the columns of our BCM (Bienenstock, Cooper, and Munro 1982)
A matrix) that have underlying causes, u (or s), and learning rule. Following up, Law and Cooper
364 Anthony J. Bell and Terrence J. Sejnowski
(1994) and Shouval (1995) used the BCM rule to If we were to accept WI as a primitive model of
self-organize oriented and somewhat localized re- the retinocortical transformation, then several objec-
ceptive fields on an ensemble of natural images. tions might arise. One might object to the represen-
The BCM rule is a nonlinear Hebbian/anti- tation learned by the algorithm: the filters in figure
Hebbian mechanism. The nonlinearity undoubtedly 19.3 are predominantly of high spatial frequency,
contributes higher-order statistical information, but even though spatial frequencies have been found to
it is less clear than in Olshausen’s network or our own spread over several octaves in cortex (Hubel and
how the nonlinearity contributes to the solution. Wiesel 1974). The reason there are so many high
Another principle, predictability minimization, spatial frequency filters is because they are smaller,
has also been brought to bear on the problem by therefore more are required to ‘‘tile’’ the 12 12
Schmidhuber, Eldracher, and Foltin (1996). This pixel array of the filter. However, active control
approach attempts to ensure independence of one of fovea-based eye movements and the topographic
output from the others by moving its receptive field nature of V1 spatial maps means that visual cortex
away from what is predictable (using a nonlinear samples images in a very di¤erent way from our
‘‘lateral’’ network) from the outputs of the others. random, spatially unordered sampling of 12 12
Finally, Harpur and Prager (1996) have formalized pixel patches. Changing our model to make it more
an inhibitory feedback network that also learns realistic in these two respects could produce di¤erent
nonorthogonal oriented receptive fields. results.
Another important issue with regard to redun-
dancy reduction is the significant redundancy across
19.4 Biological Significance the encodings of neighboring image patches. The
spatial decorrelation of natural images in a wavelet
The simplest properties of classical V1 simple cell representation leads to suppressive interactions be-
receptive fields (Hubel and Wiesel 1968) are that
tween filters in neighboring patches (Schwartz and
they are local and oriented. These are properties of Simoncelli 1999), similar to what has been reported
the filters in figure 19.4, while failing to emerge in the primary visual cortex (Das and Gilbert 1999).
(without external constraints) in many previous self- The approach taken here can also be extended to
organizing network models (Linsker 1988; Miller redundancy that occurs in sequences of images (van
1988; Atick and Redlich 1993; Troyer et al. 1999). Hateren and Ruderman 1998). Here the inputs are
However, the transformation from retina to V1, from three-dimensional spatiotemporal patterns and the
analog photoreceptor signals to spike-coding pyra- filters have the properties of directionally selective
midal cells, is clearly much more complex than the simple cells found in the primary visual cortex.
WI matrix, with which we have been working. The properties of neurons in the visual cortex
Nonetheless, evidence supports a feedforward depend on experience as well as genetically deter-
origin for the oriented properties of simple cells mined mechanisms, so it is natural to ask whether
in the cat (Ferster et al. 1996). Also the ZCA filters there are biological ways that an ICA algorithm could
approximate the static response properties of gan- be implemented. Although the learning rule we used
glion cells in the retina and relay cells in the lateral is nonlocal, it involves a feedback of information
geniculate nucleus, which, to a first approximation, from, or within, the output layer. There are many
prewhiten inputs reaching the cortex.
Learning to Find Independent Components 365
ways that such a biophysical self-organizational pro- type of object (i.e., orientation). Complex cells in
cesses could be accomplished using local spatial media area V1, which are somewhat location invariant, and
where the feedforward and the feedback of informa- neurons further up the visual processing pathways,
tion are tightly functionally coupled (Bell 1992; which have many invariant properties, present a
Eagleman et al. 2001). huge challenge to unsupervised learning models.
Regardless of whether any biological system Can their properties be predicted (or retrodicted)
implements an unsupervised learning rule such as and their coding properties thus explained?
ICA, the results allow us to interpret the response We believe the answer to this question is yes, and
properties of simple cells in visual cortex as a form of that it will involve the formulation of algorithms
redundancy reduction, as Barlow conjectured. Care related to ICA, in which group-theoretic symmetries
must be taken, however, in drawing strong con- in probability distributions are identified with the
clusions about visual cortical encodings, from models subspaces in which they are embedded. Von der
consisting of only a single static linear transformation. Malsburg has argued convincingly for many years
that invariant coding and ‘‘feature binding’’ are the
same problem, so we expect such learning algo-
19.5 Conclusion rithms will help bridge, in an information-theoretic
way, the di‰cult gap between sensory and percep-
What coding principles predict the formation of
tual learning.
localized, oriented receptive fields? Barlow’s answer
This will also greatly increase the computational
was that edges are suspicious coincidences in an im-
power of abstract unsupervised learning techniques.
age. Based on the principles of information theory
(Cover and Thomas 1991), Barlow proposed that
our visual cortical feature detectors might be the end
result of a redundancy reduction process (Barlow
1989; Atick 1992), in which the activation of each
feature detector is as statistically independent from the
others as possible.
We approached this problem through unsuper-
vised learning in a single layer of linear filters based
on an ensemble of natural images. The localized edge
detectors that were produced have phase sensitivity
as a result of the sensitivity of ICA to higher-order
statistics.
Edges (or rather, areas of local contrast) are the
first level of structure in images, being detectable by
linear filters alone. The analogous cells in area V1,
called ‘‘simple cells,’’ are the last in the visual system
to fit a ‘‘cardinal cell’’ model (von der Malsburg
1999)—that is, there is one cell for each location and
This page intentionally left blank
Top-Down Information and Models of
Perceptual Learning
Michael H. Herzog and Manfred Fahle
20
Abstract models (a mixture of supervised and unsupervised
learning) using stimuli that the system can classify
Current neural network models, mostly concerned with internally have also been suggested (see Weiss,
bottom-up processes, such as finding optimal parameters Edelman, and Fahle 1993; Fahle and Edelman 1993).
for a given set of data (which correspond to the stimuli of
Simulations with so-called radial basis function
experiments), do not incorporate top-down information,
such as preselecting features or internal knowledge. New (RBF) networks have replicated the e¤ect of orien-
experimental results, however, show that attention and tation specificity for vernier stimuli (Poggio, Fahle,
other higher cortical processes play an important role in and Edelman 1992), o¤ering a possible explanation
perceptual learning issues. After briefly reviewing current for the process underlying perceptual learning: the
(mathematical) learning models, we present these new
stimuli were used to synthesize highly specific centers
results and sketch out a framework of perceptual learning
that takes top-down influences into account. used as the bases for interpolation (see Sinha and
Poggio, chapter 15, this volume). Because finding
these centers is task specific, no transfer of improve-
ment occurs even between similar tasks. In this
20.1 Introduction chapter, we show that in addition to these purely
feedforward mechanisms top-down aspects such as
One predominant feature of perceptual learning is attention are also important for perceptual learning.
its specifity for certain stimulus dimensions such as Because we will often refer to experiments using a
orientation, spatial frequency, direction of motion, vernier discrimination task, let us briefly describe this
retinal position, and the eye of presentation (e.g., paradigm (for a more detailed description, see chap-
Fiorentini and Berardi 1980; Karni and Sagi 1991; ter 11, this volume). A vernier consists of two almost
Shiu and Pashler 1992; Fahle, Edelman, and Poggio aligned straight bars of the same orientation that are
1995; Schoups, Vogels, and Orban 1995; Ahissar and slightly displaced relative to each other by an o¤set
Hochstein 1997; Crist et al. 1997; Rivest, Boutet, that might be much smaller than the smallest diam-
and Intriligator 1997; but see also Liu 1999; Liu and eter of a retinal photoreceptor. Despite the small size
Weinshall 2000; chapters 9–12). Models of percep- of this spatial o¤set, most observers are able to dis-
tual learning have been mainly concerned with criminate its direction. More importantly, most
describing these specificities and mainly focused on subjects improve this ability with practice. The phe-
the isolated set of data (figure 20.1). Both supervised nomenon of spatial resolution below the diameter of
and unsupervised learning schemes show good agree- a retinal photoreceptor is called ‘‘hyperacuity.’’ It is
ment with the data (see Moses, Schechtman, and also possible to define hyperacuity tasks with three-
Ullman 1990; Poggio, Fahle, and Edelman 1992; dot verniers (figure 20.2, panel a). It is believed that
Sundareswaran and Vaina 1994). Bootstrapping changes on the very early stages of visual processing
368 Michael H. Herzog and Manfred Fahle
Figure 20.2
Figure 20.1
(a) Three-dot vernier, whose middle point is shifted
Three-layer neural network with a single output unit, the
slightly to the left or right of the imaginary line through
most common architecture for supervised learning. Here
the outer points. (b) Nine-dot stimulus, for which 84
the output
P 4 for a particular P 4input P vector x ¼ ðx1 ; x2 Þ is
three-dot spatial tasks can be defined, including diagonals,
FðxÞ ¼ i¼1 ai gðx wi Þ ¼ i¼1 2
ai gð j¼1 xj wij Þ, where ai
‘‘triangles,’’ and the like. Additional tasks can be defined
represents the ‘‘synaptic’’ weights between the hidden units
with more or less than three points; the possible number of
and the output layer unit, and wij represents the weights
tasks grows exponentially with the number of (homoge-
between the two input units and the four hidden layer
neous) features of the stimuli (dots). (c) Features correlated
units. Not all ‘‘synaptic’’ weights are indicated.
to the target dots. If the middle dot is o¤set to the right,
the bars are always rotated clockwise; if the middle dot
in the cortex are involved in the learning process is o¤set to the left, the bars are always oriented counter-
clockwise. (d ) Schematic graph of performance for se-
because improvement through learning is specific
quential training of two possible tasks, with percentages of
both for the orientation and the eye used during correct responses plotted against time (or block numbers).
training (e.g., chapter 11, this volume; Gilbert 1994; (e) Vernier discrimination might use the two vertical seg-
but see also Mollon and Danilova 1996). Neurons ments on task 1 and the two horizontal segments on task 2.
that are monocularly driven and orientation selective Throughout the whole experiment, vertical and horizontal
are mostly found in the primary visual cortex, area verniers are presented simultaneously. Only attention to
the particular task is di¤erent. If observers attend to task 1
V1 (Hubel and Wiesel 1959), which proves to be
first, an improvement of performance is found that does
more plastic even in adult animals than previously not transfer to task 2. Thus passive viewing does not yield
thought (see Eysel, chapter 3, this volume; Gilbert learning (cf. chapter 11.13).
1994).
taining s samples. Models for which classification of with the weight vector wi ¼ ðwi1 ; . . . ; wir Þ of unit i
stimuli is a necessary component are called ‘‘super- and the activation vector y ¼ ð y1 ; . . . ; yr Þ of the
vised’’; those for which it is not, ‘‘unsupervised.’’ The preceding layer (containing r units). The scalar y wi
dependency on feedback is one of the major features is fed into an activation function g. Popular activa-
used to characterize learning models. tion functions are sigmoid functions such as the
For the sake of simplicity, we focus mainly on 1
common sb ðxÞ ¼ .
feedforward networks, where information propa- 1 þ e2bx
gates in only one direction and neurons belonging to
the same layer are not connected with each other. 20.2.1 Supervised Learning with a Teacher
Models where all neurons are interconnected en-
counter problems that do not arise in feedforward Supervised learning means adjusting the ‘‘synaptic’’
networks. Recurrent models, for example, have to weights to obtain the desired output d k for every
reach a ‘‘convergent’’ state to classify a given stimulus. given x k of a set of training data X ¼ fx 1 ; . . . ; x s g.
In Hopfield networks, this state is reached when the The goal is to minimize the norm kd k Fðx k Þk for
reciprocal weights of two neurons are identical, all pairs of data values and desired outputs ðx k ; d k Þ.
whereas, in cortical networks, the reciprocal con- The adjustment can be achieved with a learning rule
nections between two neurons, if they exist at all, written as
will rarely by of equal strength. It should also be
mentioned that, in the brain, neither the assumptions wijtþ1 ¼ wijt þ aLðG t ; x k ; ðd k Fðx k ÞÞÞ; ð20:2Þ
of feedforward networks of full connectivity nor of where G t is the set of all weights of the neural net-
lack of connectivity inside a layer are met. For ex- work at time t. The vectors x k and G t completely
ample, in the primary visual cortex most, but not all, determine the state of the net at time t. The teacher
connections of a neuron terminate on other primary term ðd k Fðx k ÞÞ indicates the error the network
visual cortex neurons (Peters, Payne, and Rudd 1994). produced at time t after the presentation of the vec-
The relevance for modeling of the experimental tor x k . This procedure is called ‘‘data labeling’’ be-
results presented here is largely independent of the cause the teacher is attaching a label d k to each data
connectivity of the architecture. point x k . In the case of no error, LðG t ; x k ; 0Þ ¼ 0;
A feedforward neural network (figure 20.1) usu- thus no modifications occur. The learning rate a
ally consists of a number of layers, each composed determines the speed of the learning process, and the
of a defined number of units. Mathematically, such function L specifies the di¤erent on-line techniques
networks can be described as a function F : Rn ! such as backpropagation. In supervised learning with
Rm , where n is the number of input units. Their a teacher, two components are necessary: (1) a de-
activation is the input vector x ¼ ðx1 ; . . . ; xn Þ and sired output has to be given for every stimulus pre-
m is the number of output units. The output of sentation; and (2) the di¤erence between the actual
any particular unit i is calculated as the dot product and the desired output has to be evaluated. It is im-
y wi , where portant to note that not every feedback signal can
X
r serve as a teacher signal (see, for example, the
y wi ¼ yj wij ; ð20:1Þ experiment providing block feedback in section
j¼1
370 Michael H. Herzog and Manfred Fahle
20.3.2). Supervised models with a teacher do not Indeed, none of these models was ever proposed
allow learning without a data labeling mechanism. for this purpose; thus we do not discuss them here.
Moreover, most of the problems of supervised learn-
20.2.2 Unsupervised Learning ing models, as discussed below, apply to these archi-
tectures as well.
In unsupervised learning models, only the data X ¼
fx 1 ; . . . ; x s g are given; the model learns by extract- 20.2.4 Summary
ing ‘‘features’’ just from the data. Well-known ex-
amples are competitive or Hebbian learning rules. In all models described above, learning exclusively
Most of these models describe learning as a system of focuses on the set of data. Neither preprocessing,
exposure-dependent rules adjusting the strengths of such as a selection of interesting coordinates of the
connections (weights) between their elements strictly input vectors, nor any other top-down information
according to the stimulus. These procedures are is incorporated. All modifications made with respect
totally independent of any top-down e¤ects. As in to the stimuli presented are implicit. For example,
supervised learning models, the input vector prop- the probability distribution of the stimuli is implicitly
agates through the network, and the output of a built into the synaptic weights but is neither esti-
particular unit is determined in the same way. mated explicitly nor stored in an independent mem-
However, no teacher term ðd k Fðx k ÞÞ is present in ory that may control the learning process. The same
the learning rule, which may be written as holds true for the rate of feedback signals. A bias of
feedback signals favoring one decision class over an-
wijtþ1 ¼ wijt þ aLðG t ; x k Þ: ð20:3Þ
other is not explicitly represented, thus cannot in-
To give an example: during the experiments, owing fluence the learning process directly. The learning
to noise (e.g., eye tremor), the repeated presentation rate a in equations 20.2 and 20.3 above is a scalar,
of stimuli may form two classes of inputs for the two whereas, in most models, a has to approach zero to
fixed vernier o¤sets (and not only two vectors). A ensure convergence. If a does not decrease, the out-
winner-take-all mechanism might find the appro- put of the network may di¤er widely when the same
priate weights to separate these classes. Purely expo- stimulus is presented at di¤erent times. Most learning
sure-dependent unsupervised learning rules clearly rules thus use a t with a t ! 0, which is to say that,
allow learning without feedback; they predict that with time, the system loses its plasticity. Moreover,
learning will be independent of the particular feed- the amount of modification achieved by a single
back condition. presentation of an input vector may be quite limited.
feedback; and (3) internal criteria and information. Meinhardt and Grabbe (in press) for spatial fre-
We present new experimental findings on all three. quency discrimination. When Herzog and Fahle
(1994) presented a vertical and a horizontal (line)
20.3.1 Selection Processes vernier simultaneously, they found only partial
transfer between the stimuli presented at di¤erent
Most neural networks treat the stimulus as a whole, orientations (chapters 11–13). This result shows that
incorporating neither a task-guided selection of fea- attention can actively select between the learning
tures of the stimuli nor any other top-down infor- even of very similar tasks and between neighboring
mation. In contrast, general (external) knowledge stimuli. In a neurophysiological study, Ahissar et al.
about particular features may play an important role (1992) found that monkeys trained on an auditory
for perceptual learning. What happens, for example, task learned only the behaviorally relevant features.
if two (or more) tasks can be defined for a set of Taken together, these results suggest that selection
stimuli (see figure 20.2)? Are all tasks learned or only processes based on general knowledge about the set of
the instructed ones? If only the instructed tasks, a stimuli as a whole (e.g., What are the relevant fea-
selection of particular features of the set of stimuli tures?) play an important role in perceptual learning.
has to occur, which can be considered to involve
attention. The typical experiment to test whether Why Is the Brain Not Learning All Possible
attention is involved in a learning task is to present a Tasks?
set of stimuli for which at least two tasks can be A stimulus configuration does not determine the
defined (see figure 20.2). Observers train on the first task: many tasks can be defined for almost all sets of
task in the first sessions and on the second in suc- stimuli, and sheer complexity, makes it impossible to
ceeding sessions; the amount of improvement trans- learn all tasks. In figure 20.2, panel b, 84 di¤erent
ferred from the first to the second task is determined. three-dot tasks might be defined. In principle, the
Note that the set of stimuli is always the same; only number of tasks grows exponentially with the num-
the tasks di¤er. ber of features in the stimulus: to every subset of fea-
Shiu and Pashler (1992) presented lines di¤ering tures (i.e., the dots in panel b), a task can be defined.
in both orientation and luminance. Training on the Because no knowledge about the relevant stimulus
luminance discrimination task did not improve per- features can be incorporated and because a given task
formance on the orientation discrimination task af- can be performed with almost infinitely many sets of
terward. This result proves that learning does not stimuli, this problem is a major challenge for unsu-
transfer between two ‘‘visual dimensions.’’ A similar pervised learning models. Consider the three-dot
finding was made by Boutet, Intriligator, and Rivest vernier discrimination task in figure 20.2, panel a.
(1995) on a color versus motion task. Investigating The same task can be performed with the stimuli
the transfer from a globally to a locally defined tex- displayed in panel b, which have six additional
ture discrimination task and vice versa, Ahissar and neighboring dots; although the task remains the same,
Hochstein (1993) found no transfer from the global the set of input vectors has changed dramatically.
to the local task, and only a small partial transfer vice Where the stimuli are correlated, the problem
versa, indicating that attention was involved in spatial might be even harder. Additional features correlated
discrimination tasks. Similar results were obtained by with the target stimuli might be used to solve the
372 Michael H. Herzog and Manfred Fahle
task—especially if these features have a larger d 0 , that How do humans use external feedback? A pio-
is, if they are ‘‘more easily’’ discriminated than the neering study on the dependency of feedback for
‘‘original’’ features. Figure 20.2, panel c, illustrates perceptual learning was conducted by Shiu and
such correlated features. The two targets to be dis- Pashler (1992) using an orientation discrimination
criminated from each other consist of a dot vernier, paradigm. Here we discuss data from a more recent,
with the middle dot o¤set to the left or right. If the larger study (Herzog and Fahle 1997) that used a
middle dot is o¤set to the right, the bars are always vernier discrimination task to investigate perceptual
rotated clockwise and if the middle dot is o¤set to learning. Vernier stimuli were presented under dif-
the left the bars are always rotated counterclockwise. ferent but comparable feedback conditions:
If the vernier discrimination task is ‘‘more di‰cult’’
1. Correlated trial-by-trial feedback. An incorrect re-
than the orientation discrimination task defined by
sponse was followed immediately by an acoustical
the correlated bars, the latter task might be ‘‘learned’’
error signal, whereas no tone occurred after a correct
instead of the vernier discrimination task. In super-
response.
vised learning models, weights may be changed ac-
cording to the correlated features because feedback is 2. No feedback.
also correlated to them. After decorrelating the tasks, 3. Block feedback. At the end of each block of eighty
a deterioration of performance will result. stimuli, a score of correct responses was displayed;
Thus the set of stimuli does not determine the during a block, no error signals were provided.
task, nor vice versa. Because not all tasks can be 4. Uncorrelated trial-by-trial feedback. All responses
learned, it is important to incorporate knowledge were labeled as incorrect with a probability of 0.5.
about the features relevant for the task. Attention-
like mechanisms might contribute to the solution of Correlated trial-by-trial or block feedback (condition
the problem of selecting features, which becomes 1 or 3) improved the speed of learning as well as
more prominent if di¤erent tasks have to be learned overall performance, whereas uncorrelated trial-by-
in succession (see also section 20.3.3). trial or no feedback (condition 2 or 4) slowed down
or even abolished improvement of performance (see
20.3.2 External Information also Ball and Sekuler 1987; Shiu and Pashler 1992;
chapters 11, 12). With the setup described above, we
Role of External Feedback did not find a significant improvement of perfor-
One of the main characteristics used to classify neural mance if no feedback was provided, although some
networks is their dependency on external feedback, subjects improved their performance under this con-
which can be viewed as additional external infor- dition, and in similar experiments, observers improved
mation (about ‘‘correct’’ classifications) for each par- significantly even in the absence of error feedback
ticular stimulus. Because, as noted above, supervised (e.g., McKee and Westheimer 1978; for long-term
neural networks with a teacher have to rely on ex- learning, see Shiu and Pashler 1992; chapters 11, 12).
ternal feedback to label the data, they propose that Thus learning without external feedback is possible;
perceptual learning without external feedback is im- whereas the improvement of performance is slower
possible, whereas ( pure) unsupervised learning mod- without than with feedback, the positive e¤ect of
els propose quite the contrary, and are indi¤erent to feedback is neither very specific nor accurate: reduced
whether or not feedback is provided. feedback in the block feedback condition does not
Top-Down Information and Models of Learning 373
change the results dramatically. On the other hand, propriate connections for the task must be found
manipulated feedback abolishes learning. Because before adjusting the appropriate sizes of the weights,
unsupervised learning models incorporate no feed- which may in part explain why the experimental
back-dependent computations, they cannot explain results do not supply any positive evidence for the
this graded dependency on external feedback. And existence of a supervised procedure with a teacher in
because external feedback is not used to label the data perceptual learning.
and to compute a teacher signal, supervised learning
models with an external teacher cannot be used to Learning with Insight
explain the e¤ects of feedback on perceptual learning If after a period of nonimprovement with stimuli
either. very di‰cult to discriminate (small d 0 ), stimuli with
a large d 0 are briefly presented, an immediate and
Why Is External Feedback Not Used to Label strong improvement may follow, the ‘‘Eureka’’ e¤ect
the Data? (Ahissar and Hochstein 1995; Rubin, Nakayama,
The average number of connections of a neuron in and Shapley 1997; Rubin, Nakayama, and Shapley,
the monkey neocortex is about 4 10 3 to 10 4 syn- chapter 13, this volume; Meinhardt 2001). Because,
apses (Beaulieu et al. 1992), creating networks with however, the amount of modification after each
an incredibly large number of weights to be adjusted. stimulus is determined by the learning rate and thus
As an example, imagine that an output neuron is may be quite limited, many learning models cannot
coding for a particular task in a three-layer feedfor- reflect this behavior by incorporating the additional
ward network and a supervised learning algorithm crucial information. Moreover, in supervised mod-
tries to backpropagate the error. This output neuron els, correctly classified stimuli that result from a large
is connected to 4 10 3 neurons in the hidden layer d 0 , corresponding to a zero error d FðxÞ ¼ 0 (see
and each of the 4 10 3 neurons has 4 10 3 con- equation 20.2), do not yield any modification.
nections to the input layer. In total, at least 1:6 10 7 Therefore, information already ‘‘known’’ by the sys-
weights are involved in this network to improve tem does not contribute to the learning process.
only one particular task. But the situation in the
brain is even more complex. Not all connections 20.3.3 Internal Criteria and Information
projecting to a neuron in the cortex originate ex-
clusively from areas ‘‘below,’’ nor are all related to The last two subsections described the influence of
the particular task. Connections originate from non- external top-down information on perceptual learn-
cortical structures, from neurons in the same area, and ing. Here we present results of experiments show-
from other areas. Less than 10% of the inputs to neu- ing that the brain builds up and uses internal criteria
rons in area V1 originate from the lateral geniculate and information to actively control the learning
nucleus (Peters, Payne, and Rudd 1994). An error process.
backpropagation procedure might change some of In experiments providing biased feedback, Herzog
these weights, at the same time erasing weights and Fahle (1999) found their subjects used internal
adjusted for di¤erent tasks. Because so many synapses criteria about the correctness of responses. They
need to be modified (a quantitative problem) and provided manipulated feedback on a binary choice
because a mechanism is needed to prevent adjusting task using vernier stimuli with di¤erent o¤set sizes,
the ‘‘wrong’’ ones (a qualitative problem), the ap- where one of the stimuli was labeled as belonging to
374 Michael H. Herzog and Manfred Fahle
that simply extending the existing models with some Neurons in higher cortical areas have to send task-
additional ‘‘independent’’ structures will not improve related information to the appropriate lower level
them: top-down information must be incorporated areas to overcome the various selection problems.
in an integral way. Because many tasks have to be performed using the
It should be noted that the components of neural same neurons, we suggest that information is gated
network models were criticized (Shepherd 1990; in a task-dependent way through a processing stage.
Gardner 1993). One problem here is that many as- In sketching out our framework for perceptual
pects of neurons, like their compartment structure learning, we propose that models may be divided
or (nonlinear) intersynaptic computations, are over- into three parts: the first to consist of conventional
simplified or not modeled at all. The importance of feedforward propagation of the stimulus; the second
these aspects for a biologically realistic description of to evaluate any available feedback; and the third to
neural processing cannot be judged yet. Moreover, control the propagation of information, using recur-
most neurons are either excitatory or inhibitory, but rent operations to change the gating properties of
not both, implying that they cannot change their processing units. Because of the huge number of
synaptic weights from positive to negative values and synaptic connections, neurons, and modules, selec-
vice versa (Dale’s law; Shepherd 1990; Kandel and tion processes must be involved that modify just a few
Schwartz 1991). of the many combinations of the processing units,
In summary, a serious problem in modeling per- depending on the chosen task (figure 20.4). In our
ceptual learning consists in how remote information framework, internal and (if provided) external feed-
related to the task and stored in higher cortical areas back evaluates the ‘‘success’’ of this choice and con-
can control learning in the very early processing trols the temporal application of the update (i.e.,
stages. Only these early levels contain neurons with learning) mechanism. Depending on the result, the
small receptive fields that might provide the infor- next element in the selection process is chosen ac-
mation necessary to process the selected task (see cording to feedback evaluation, a priori knowledge,
section 20.1; and chapter 11, this volume). Unsu- and other top-down influences (figure 20.4).
pervised learning rules use purely local information, Incorporating selection mechanisms can solve the
but cannot integrate top-down knowledge about the problem of choosing relevant features. If no exact
required task. Supervised models, on the other hand, information about the ‘‘best’’ choice exists, di¤erent
can incorporate general knowledge but cannot real- choices may be evaluated successively. In this sce-
ize the local changes necessary for perceptual learn- nario, it does not matter whether di¤erent tasks
ing while obeying biological constraints. are learned simultaneously or in succession, thus
bypassing the plasticity-stability dilemma. The se-
lection mechanisms can also solve the problem of
20.5 Toward a Framework Incorporating choosing which of the many weights must be
Top-Down Information modified. In this framework, feedback is not used to
calculate a teacher term that depends on the known
It follows from the discussion above that a model of classification of stimuli. Internal and external feed-
perceptual learning has to use feedback connections, back evaluates performance, thus actively controlling
possibly of a cholinergic nature (see chapter 2.1.7). the learning process via the learning rate. Learning
Top-Down Information and Models of Learning 377
Figure 20.4
Major components of a recurrent model involving top-down influences and a feedback-controlled learning rate a. A stim-
ulus is presented and, depending on a selection mechanism, its representation is gated through one (or more) processing
layers to a decision stage. The selection mechanism chooses particular processing units and modifies their properties to
achieve a ‘‘correct’’ classification of the stimulus. Processing units may be single neurons, groups of neurons, whole modules,
or just axons of neurons. Based on the gating, a decision is made about the stimulus resulting in a motor response (not shown
in this figure); this decision creates an internal feedback term, which depends on the discriminability, i.e., determined by
both the physical properties of stimuli and the properties of gating. The value of the feedback term is evaluated, together
with external feedback (if provided), and fed into the selection mechanism, which is also controlled by top-down mecha-
nisms such as attention. Each time a particular task is carried out, a corresponding selection action is chosen. The goal of
learning is to find ‘‘good’’ actions to discriminate between the features that are ‘‘important’’ for the task. However, it is not
necessary to design a special detector for this task as long as the features are discriminated separately. In the framework
depicted here, two mechanisms are involved: a selection mechanism, b, to control which of the units are chosen (in this
figure the black squares) and an update mechanism, a, to control the processing changes of the units chosen by b. Both
mechanisms interact with each other. The control of a depends heavily on internal and external feedback. For example, b
selects a group of neurons and a Hebb-rule is applied to change the synaptic weights of these neurons. Feedback controls the
speed of the update process by controlling the learning rate a. Modifications in the top-down parts are also allowed. The
major di¤erence from the neural network models reviewed in this chapter is that feedback actively controls the learning
process in a closed loop with an (implicit) feedback memory. Additional top-down aspects are easily incorporated.
without external feedback is possible because inter- and internal criteria (such as a priori assumptions
nal feedback is employed, although learning without about the statistics of stimuli), learning via the mech-
external feedback is slower because the update pro- anism a is suppressed. Learning by insight and exter-
cess cannot be accelerated. Learning by insight may nal feedback may be related. Stimuli that are more
help make good choices: easily discriminated stimuli easily discriminated can guide learning of stimuli that
‘‘suggest’’ which modules are ‘‘important’’ for the are harder to discriminate by providing internal feed-
particular task, thus facilitating the learning process. back about correctness of responses (see above). Thus
If, in comparison, there are disturbances in the pro- providing ‘‘easy’’ and ‘‘hard’’ stimuli might be inter-
cessing of stimuli (such as those mediated by feedback) changeable with providing only ‘‘hard’’ stimuli and
378 Michael H. Herzog and Manfred Fahle
external feedback. (Here it should be noted that the the di¤erent vernier tasks is believed to occur. Fahle
amount of learning depends on the individual base- (1997) also found no transfer between discrimination
lines of subjects. Faster improvement is found for tasks employing verniers, chevrons, and lines with
subjects training with a vernier with a larger o¤set di¤erent orientations, all three stimuli believed to be
compared to observers training with a smaller o¤set processed by orientation sensitive cells in V1. A long-
size—even the starting performance levels of the lasting modification of receptive fields according to
subjects, measured in percentages of correct re- one of these tasks should modify performance on the
sponses, are virtually identical; see also Fahle and other tasks. However, this proposition could not be
Henke-Fahle 1996.) confirmed. We suggest that receptive fields are
In our framework, components are integrally in- modified via gating but only as long as the task is
terconnected. Models may take into account the performed to avoid any interference with other tasks.
functional di¤erences between excitatory and inhib-
itory systems, as is known from several anatomical
20.6 Conclusions
and electrophysiological investigations (e.g., Shep-
herd 1990; Kandel and Schwartz 1991). The feed-
In the last ten years, many pioneering studies have
forward pathway may be modeled as the excitatory
discovered mechanisms of perceptual learning; many
part, with feedback loops assumed to exert recurrent
models have reproduced the results of these studies,
inhibition on the feedforward excitation. (For a con-
largely by focusing on the stimuli themselves. More
crete example of modeling learning in a vernier
recent experimental results suggest, however, that
discrimination task, see Herzog and Fahle (1998).
Although many other models are conceivable; see, the stimuli are only one part of the story. Other
aspects, as yet not included in models, are important
for example, Williamson 1999).
An important feature of our framework is the as well. In this chapter, we showed that the role of
feedback for the learning process was misunderstood,
volatility of the proposed gating operations, which
and the role of selection and control processes
are enabled only when the corresponding task is
underestimated.
performed. Thus no long-lasting neurophysiological
It seems that the learning of even simple stimuli,
modifications, such as a permanent change of tuning
such as those used in most experiments on percep-
curves of neurons, are expected. This feature agrees
tual learning, cannot be tackled by simple models.
well with recent psychophysical and neurophysio-
Rather, perceptual learning requires highly sophisti-
logical findings. Schoups (chapter 5, this volume) did
cated architectures. The experimental results suggest
not find any changes in the responses of area V1
neurons after learning a perceptual orientation dis- that most learning phenomena cannot be tracked
down, and that a single, comprehensive model of
crimination task. Fahle (chapter 11, this volume)
perceptual learning thus cannot be fully specified.
showed that no transfer of learning occurs between
We may never find the basic elements of learning
discrimination tasks employing verniers with a dif-
that might be used as an ‘‘alphabet’’ for learning
ference of orientation as small as 10 , a surprising
more complex tasks because these elements work
result because the half-width of receptive fields of
together. On the other hand, because top-down
V1 neurons is believed to be 30 (Movshon and
influences and other important aspects play an im-
Blakemore 1973) and thus an interaction between
Top-Down Information and Models of Learning 379
Acknowledgments
cytoarchitectonic areas Distinct regions of the GAD Glutamate decarboxylase. Enzyme for the
neocortical mantle identified by di¤erences in cell production of GABA from glutamate.
size, packing density, and laminar arrangement.
generalization The ability of a learning system to
dark-rearing Raising of animals in complete classify novel patterns (e.g., to respond adequately
darkness (in a dark room). to a novel stimulus), to which it has not been pre-
distal Farther away from a point of reference (op- viously exposed.
posite of proximal ). generator Entity or process underlying stochastic
EPSP Excitatory postsynaptic potential. Depola- data that is responsible for the regularities existing in
rizing event at the postsynaptic membrane of a syn- a data set. Typically, the underlying regularities are
apse caused by the action of a transmitter substance obscured by measurement imperfections, extraneous
at ligand-gated (fast) or metabotropic receptor– factors, and noise, and must be uncovered, for ex-
coupled (slow) channels that are permeable for so- ample, through learning.
dium and potassium. Increases the probability of glutamate Most important excitatory transmitter
reaching the firing threshold. substance in the brain. Acts on NMDA and non-
firing threshold Membrane potential level (nor- NMDA (AMPA, kainate) as well as metabotropic
mally in the range of 50 to 60 mV) that has to glutamate receptors.
be reached to elicit a regenerative sodium action gray matter General term for regions of the cen-
potential. tral nervous system rich in neuronal cell bodies and
GABA Gamma-aminobutyric acid. Most impor- neuropil, characteristically gray in outward appear-
tant inhibitory transmitter substance in the brain. ance; includes the cerebral and cerebellar cortices,
the nuclei of the brain, and the central portion of the
GABAA Receptor for GABA coupled to a ligand- spinal cord. See also white matter.
gated channel permeable for chloride. Action: fast
inhibition due to hyperpolarization (to about 70 Hebbian learning rule Postulated by Donald O.
mV) and shunting. Hebb in 1949, this rules states that a synapse is
strengthened when its pre- and postsynaptic elements
GABAB Metabotropic receptor for GABA that are synchronously activated. Theoretical background
activates a G-protein-coupled channel permeable for of long-term potentiation (LTP).
potassium. Action: slow inhibition due to hyper-
hippocampus Cortical structure in the medial
polarization (up to about 90 mV) and inhibitory
postsynaptic potential (IPSP). portion of the temporal lobe; in humans, associated
with short-term declarative memory, among many
Gabor patch Circular striped pattern (alternating other functions.
dark gray/light gray) that is generated by modulating
a sine or cosine wave grating with a two-dimensional hyperexcitability Activity state above normal.
Can be due to reduced firing threshold, depolariza-
Gaussian envelope. Gabor patches selectively excite
cells tuned to a specific orientation, spatial frequency, tion of a cell, or increased transmitter levels.
and spatial location.
Glossary 383
immunohistochemistry Method of labeling spe- potentiated. LTP is associative: weak synapses when
cific cells by means of an antigen-antibody reaction activated synchronously with stronger inputs to the
or an irreversible receptor agonist. same cell can be strengthened as well.
intracortical Within the cortex. magnification factor Factor describing the rela-
tion between a peripheral sensory surface and the
in vitro Literally, ‘‘in glass.’’ Outside the living
body. Brain tissue studied in vitro is cut in about 400 size of its central representation. The magnification
factor is as a rule directly related to the receptor
mm slices that are put into oxygen/carbogen bubbled
density in the periphery: the fovea of the retina, the
bath of artificial cerebrospinal fluid (aCSF). Cells can
frequencies related to speech in the cochlea, and the
be recorded extra- or intracellularly or by patch
lips and fingertips of the skin have the largest cortical
clamp methods in a stable situation under micro-
magnification factor.
scopic control.
manifold Smooth ‘‘surface’’ embedded in a
in vivo Literally, ‘‘in living.’’ Within the living
body. In vivo is the only way to study nerve cells higher-dimensional space, such as a curve drawn on
a sheet of paper and the sheet of paper itself,
under natural conditions with original sensory input
or intact motor output and in the intact system en- embedded in the three-dimensional space.
vironment. MAP-2 Microtubuli-associated protein 2. Asso-
ciated with axonal transport.
IPSP Inhibitory postsynaptic potential. Hyper-
polarizing event at the postsynaptic membrane of a measurement space If the output of each unit in
synapse. Involved ion channels are permeable for a sensory pathway is thought of as a dimension along
chloride or potassium. Reduces the probability of which the signal can vary, the entire ensemble of
reaching the firing threshold. units (e.g., the ganglion cells in the retina) is seen to
span a measurement (representation) space whose
LTD Long-term depression. Weakening of syn-
nominal dimensionality is equal to the number of
apses induced by low-frequency stimulation (causing
units.
a situation not in accord with the Hebbian learn-
ing rule). Correlated with a low intracellular MLP Multilayer perceptron. Layered neural net-
calcium level. In neurophysiology, persistent weak- work architecture in which the response of each unit
ening of synapses based on past patterns of activity. is determined by computing the inner product be-
tween its inputs and a vector of weights, and by
LTP Long-term potentiation. Long-lasting (hours
passing the result through a nonlinearity. MLPs are
to days) strengthening of synapses induced by high-
usually trained (i.e., the weights are adjusted) by
frequency stimulation (causing a situation in accord
backpropagation of errors.
with the Hebbian learning rule). Correlated with
an increased intracellular calcium level. The e¤ect is monocular deprivation Deprivation of sight in
mediated by postsynaptic mechanisms (e.g., phos- one eye. See also binocular deprivation.
phorylation of membrane receptors) as well as pre- MRI Magnetic resonance imaging. Brain imaging
synaptic mechanisms (e.g., increased transmitter
technique based on the detection of changes in an
release). LTP is specific: only the suprathreshold
artificially created magnetic field.
synapse that is stimulated with high frequency is
384 Glossary
neocortex Six-layered cortex that covers the bulk primary visual cortex Cortical area in the occi-
of the cerebral hemispheres. pital lobe of the brain (Brodmann’s area 17), first
cortical stage of visual scene analysis.
neuronal plasticity Property of the nervous sys-
tem that makes it modifiable in response to use and proximal Closer to a point of reference (opposite
disuse, damage, and disease. The natural events of of distal ).
development are often also associated with the term RBFs Radial basis functions. Method for function
plasticity. Adult central nervous system plasticity is
approximation and classification in which the target
often related to mechanisms active as well in devel- function is represented as a superposition of values of
opmental plasticity. basis functions, which in turn depend only on the
neuropil Dense tangle of axonal and dendritic distances between the input and their respective
branches, and the synapses between them, that lies centers. Also, a two-layer neural network architec-
between neuronal cell bodies in the gray matter of ture implementing this approximation method.
the brain and spinal cord. receptive field Region of the body surface where
NGF Nerve growth factor. First obtained from stimulation causes a sensory nerve cell (or axon) to
salivary glands of male mice by Rita Levi-Montalcini respond. In the visual system, region in visual space
(Nobel prize 1985). Regulates axonal outgrowth and where presentation of a stimulus causes a neuron to
maintains the survival of cells when taken up from respond.
the target cells and retrogradely transported to the
regression Task of determining the mean of a
soma. The trkA receptor is the specific receptor for
distribution of points in a data set along some of the
NGF. dimensions, conditional on their location along the
NMDA N-methyl-d-aspartate. Artificial gluta- remaining dimensions.
mate agonist of one type of glutamate receptors.
retinotopy Property of central parts of the visual
The NMDA receptor–gated channel is permeable
system. The retina is mapped in these structures
for sodium, potassium, and calcium. NMDA is
(lateral geniculate nucleus, superior colliculus, pri-
blocked at resting potential by magnesium (magne-
mary and higher visual cortices) in a topographic
sium block) and can only be activated after the
way, where neighboring regions of the retina are
magnesium block has been released by depolariza-
represented in neighboring regions in the central
tion. NMDA channels are associated with LTP.
visual pathways. The maps are nonlinear with an
NT-3 Neurotrophin 3. Another member of the overproportional representation of the central retina
family of neurotrophins and nerve growth factors (magnification factor).
that includes NGF and BDNF.
retrograde A movement or influence acting from
ocular dominance columns Segregated termi- the axonal target toward the neuronal cell body
nation patterns of thalamic inputs representing the (opposite of anterograde).
two eyes in primary visual cortex of some mamma- scotoma Blind area in the visual field. The blind
lian species.
spot of the retina is due to the optic nerve head (re-
gion free of receptors) in the eye and thus a ‘‘physi-
Glossary 385
ological scotoma.’’ Pathological scotomas can be due tracer In neurobiology, substance used to trace
to retinal as well as central damage all along the (label or visualize) nerve cell connections.
visual pathway.
transfer Process of learning to perform one task
smooth mappings Mappings under which a small and then using the acquired knowledge to perform a
change in the input results in a small change in the di¤erent task. See also generalization.
output. 2-AFC Two-alternative forced-choice. Common
squint See strabismus. experimental design in psychophysical experiments.
Presented with a target during one of two subse-
statistical inference Acquisition of information
quent stimulus presentation intervals, observers have
concerning the probability distribution of the data
to indicate whether the first or the second interval
(e.g., the examples in a learning scenario), which can
contained the target. This design eliminates subjec-
then be used in decision making (e.g., as in a Baye-
tive bias observes may have with detecting the pres-
sian framework).
ence of a target in a single stimulus interval (yes/no
strabismus Misalignment of the two eyes such response).
that normal binocular vision is compromised.
vertical meridian Vertical midline of the visual
striate cortex Primary visual cortex. Located in field.
the occipital pole of the brain, the first cortical area
Weber law When the task is to discriminate two
that receives subcortical input from the retina via the
lateral geniculate nucleus of the thalamus. stimuli of intensity I and I þ DI, the discrimination
threshold DI is proportional to the pedestal I. This
support vectors Data points lying close to the law is usually violated when stimulus intensities are
would-be decision boundary in the task of learning low (close to the detection threshold).
a discrimination curve. These points have a much
white matter General term for large axon tracts in
greater influence on the boundary location (dis-
the brain and spinal cord, which characteristically
criminant curve) than others, and can be used to
have a whitish cast when viewed in freshly cut cross
speed up learning and make it more e‰cient.
section. See also gray matter.
synapsin Substance associated with synapse for-
mation in development and plasticity.
tangential connections Fiber pathways in the
cerebral cortex that extend in a direction parallel to
the cortical surface (also termed horizontal or intra-
laminar), in contrast to vertical connections that run
perpendicular to the cortical surface, from layer to
layer.
threshold Stimulus intensity at which an observer
detects a stimulus (or is able to discriminate it from
another stimulus) with a defined accuracy level, most
commonly 75%.
This page intentionally left blank
References
Abbott, L. F., Rolls, E. T., and Tovee, M. J. (1996). Ahissar, M., and Hochstein, S. (1995). How early is early
Representational capacity of face coding in monkeys. vision? Evidence from perceptual learning. In T. V. Papa-
Cereb. Cortex 6: 498–505. thomas, C. Chubb, A. Gore, and E. Kowler, eds., Early
Vision and Beyond. Cambridge, MA: MIT Press.
Abeles, M. (1991). Corticonics: Neural Circuits in the Cerebral
Cortex. Cambridge. Ahissar, M., and Hochstein, S. (1995). Time course of per-
ceptual learning probes underlying mechanisms. Perception
Adcock, N. V., and Mangan, G. L. (1970). Attention and
24: 22a.
perceptual learning. J. Gen. Psychol. 83: 247–254.
Ahissar, M., and Hochstein, S. (1996a). Perceptual learn-
Adini, Y., Sagi, D., and Tsodyks, M. (1997). Excitatory-
ing transfer over space and orientation. Invest. Ophthalmol.
inhibitory network in the visual cortex: Psychophysical
Vis. Sci. 37: 3182.
evidence. Proc. Natl. Acad. Sci. U S A 94: 10426–10431.
Ahissar, M., and Hochstein, S. (1996b). Learning pop-out
Aertsen, A. M., Gerstein, G. L., Habib, M. K., and Palm,
detection: Specificities to stimulus characteristics. Vision
G. (1989). Dynamics of neuronal firing correlation: Mod-
Res. 36: 3487–3500.
ulation of ‘‘e¤ective connectivity.’’ J. Neurophysiol. 61:
900–917. Ahissar, M., and Hochstein, S. (1997). Task di‰culty and
the specificity of perceptual learning. Nature 387: 401–
Aglioti, S., DeSouza, J. F. X., and Goodale, M. A. (1995).
406.
Size-contrast illusions deceive the eye but not the hand.
Curr. Biol. 5: 679–685. Ahissar, M., and Hochstein, S. (1998). Perceptual learning.
In V. Walsh and J. Kulikowski, eds., Perceptual Constancies,
Aglioti, S., Smania, N., Atzei, A., and Berlucchi, G. (1997).
455–498. Cambridge: Cambridge University Press.
Spatio-temporal properties of the pattern of evoked phan-
tom sensations in a left index amputee patient. Behav. Neu- Ahissar, M., and Hochstein, S. (2000). The spread of at-
rosci. 111: 867–872. tention and learning in feature search: E¤ects of target
distribution and task di‰culty. Vision Res. 40: 1349–1364.
Ahissar, E., Abeles, M., Ahissar, M., Haidarliu, S., and
Vaadia, E. (1998). Hebbian-like functional plasticity in the Ahissar, M., Laiwand, R., and Hochstein, S. (2001). At-
auditory cortex of the behaving monkey. Neuropharmacol. tentional demands following perceptual skill training. Psy-
37: 633–655. chol. Sci. 12: 57–63.
Ahissar, E., and Ahissar, M. (1994). Plasticity in auditory Ahissar, M., Laiwand, R., Kozminsky, G., and Hochstein,
cortical circuits. Curr. Opin. Neurobiol. 4: 580–587. S. (1998). Learning pop-out detection: Building repre-
sentations for conflicting target-distractor relationships.
Ahissar, E., Vaadia, E., Ahissar, M., Bergman, H., Arieli,
Vision Res. 38: 3095–3107.
A., and Abeles, M. (1992). Dependence of cortical plas-
ticity on correlated activity of single neurons and on be- Ajjanagadde, V., and Shastri, L. (1991). Rules and variables
havioral context. Science 257: 1412–1415. in neural nets. Neural Comput. 3: 121–134.
Ahissar, M., and Hochstein, S. (1993). Attentional control Albrecht, D. G., Farrar, S. B., and Hamilton, D. B. (1984).
of early perceptual learning. Proc. Natl. Acad. Sci. U S A Spatial contrast adaptation characteristics of neurones re-
90: 5718–5722. corded in the cat’s visual cortex. J. Physiol. (Lond.) 347:
713–739.
388 References
Albus, K., and Wolf, W. (1984). Early post-natal devel- Atick, J. J., and Redlich, A. N. (1993). Convergent algo-
opment of neuronal function in the kitten’s visual cortex: rithm for sensory receptive field development. Neural
A laminar analysis. J. Physiol. 348: 153–185. Comput. 5: 45–60.
Allman, J., Miezin, F., and McGuinness, E. L. (1985). Stim- Attneave, F., and Frost, R. (1969). The determination of
ulus specific responses from beyond the classical receptive perceived tridimensional orientation by minimum criteria.
field: Neurophysiological mechanisms for local-global Percept. Psychophys. 6: 391–396.
comparisons in visual neurons. Annu. Rev. Neurosci. 8:
Axelrod, S. (1959). E¤ects of Early Blindness: Performance of
407–430.
Blind and Sighted Children on Tactile and Auditory Tasks.
Alloway, K. D., and Aaron, G. B. (1996). Adaptive changes New York: American Foundation for the Blind.
in the somatotopic properties of individual thalamic neu-
Baddeley, R. (1996). Searching for filters with ‘‘interesting’’
rons immediately following microlesions in connected
output distributions: An uninteresting direction to explore?
regions of the nucleus cuneatus. Synapse 22: 1–14.
Network 7: 409–21.
Allport, D. A. (1989). Visual attention. In M. I. Posner,
Baddeley, R. (1997). The correlational structure of natural
ed., Foundations of Cognitive Science, 631–682. Cambridge
images and the calibration of spatial representations. Cogn.
MA: MIT Press.
Sci. 21(3): 351–372.
Amari, S., Cichocki, A., and Yang, H. H. (1996). A new
Badler, N., and Bajcsy, R. (1978). Three-dimensional re-
learning algorithm for blind signal separation. In Advances
presentations for computer graphics and computer vision.
in Neural Information Processing Systems, vol. 8. Cambridge,
Comput. Graph. 12: 153–160.
MA: MIT Press.
Bakin, J. S., and Weinberger, N. M. (1990). Classical con-
Angeli, A., Bruce, V., and Ellis, H. D. (1999). Getting
ditioning induces CS-specific receptive field plasticity in
familiar with faces. Perception 28, Suppl. 115.
the auditory cortex of the guinea pig. Brain Res. 536: 271–
Anton, B. S, Player, N. I., and Bennett, T. L. (1981). 286.
Transfer of perceptual learning: Role of tactual-kinesthetic
Bakin, J. S., and Weinberger, N. M. (1996). Induction of a
feedback reexamined. Percept. Mot. Skills 52: 195–202.
physiological memory in the cerebral cortex by stimulation
Arckens, L., Schweigart, G., Qu, Y., Wouters, G., Pow, of the nucleus basalis. Proc. Natl. Acad. Sci. U S A 93:
D. V., Vandesande, F., Eysel, U. T., and Orban, G. A. 11219–11224.
(2000). Cooperative changes in GABA, glutamate and
Baldassi, S., and Burr, D. C. (2000). Feature-based inte-
activity levels: The missing link in cortical plasticity. Eur. J.
gration of orientation signals in visual search. Vision Res.
Neurosci. 12: 4222–4232.
40: 1293–1300.
Artola, A., and Singer, W. (1987). Long-term potentiation
Baldi, P., and Heiligenberg, W. (1988). How sensory maps
and NMDA receptors in rat visual cortex. Nature 330:
could enhance resolution through ordered arrangements of
649–652.
broadly tuned receivers. Biol. Cybern. 59: 313–318.
Ashbridge, E., Perrett, D. I., Oram, M. W., and Jellema,
Ball, K., and Sekuler, R. (1982). A specific and enduring
T. (2000). E¤ect of image rotation or size change on ob-
improvement in visual motion discrimination. Science 218:
ject recognition: Responses of single units in the macaque
687–698.
monkey temporal cortex. Cogn. Neuropsychol. 17: 13–34.
Ball, K., and Sekuler, R. (1987). Direction-specific im-
Atick, J. J. (1992). Could information theory provide an
provement in motion discrimination. Vision Res. 27: 953–
ecological theory of sensory processing? Network 3: 213–
965.
251.
Baltes, P. B., and Lindenberger, U. (1997). Emergence of a
Atick, J. J., and Redlich, A. N. (1990). Towards a theory
powerful connection between sensory and cognitive func-
of early visual processing. Neural Comput. 2: 308–320.
References 389
tions across the adult life span: a new window to the study Baron, R. J. (1981). Mechanisms of human facial recogni-
of cognitive aging? Psychol. Aging 12: 12–21. tion. Int. J. Man-Mach. Stud. 15: 137–178.
Bar, M., and Biederman, I. (1999). Localizing the cortical Barrow, H. G., and Tenenbaum, J. M. (1981). Interpreting
region mediating visual awareness of object identity. Proc. line-drawings as three-dimensional surfaces. Artif. Intell.
Natl. Acad. Sci. U S A 96: 1790–1793. 17(1–3): 75–116.
Bara-Jimenez, W., Catalan, M. J., Hallett, M., and Gerlo¤, Bartlett, J. C., and Searcy, J. (1993). Inversion and config-
C. (1998). Abnormal somatosensory homunculus in dys- uration of faces. Cogn. Psychol. 25: 281–316.
tonia of the hand. Ann. Neurol. 44: 828–83.
Barto, A. (1989). From chemotaxis to cooperativity:
Barbas, H. (1992). Architecture and cortical connections of Abstract exercises in neuronal learning strategies. In R.
the prefrontal cortex in the rhesus monkey. In Advances Durbin, C. Miall, and G. Mitchison, eds., The Computing
in Neurology, vol. 57, ed. P. Chauvel et al., 91–115. New Neuron, 73–98. New York: Addison Wesley.
York: Raven Press.
Baskerville, K. A., Schweitzer, J. B., and Herron, P. (1997).
Barlow, H. B. (1950). The receptive fields of ganglion cells E¤ects of cholinergic depletion on experience-dependent
in the frog retina. In Proceedings of the Eighteenth Interna- plasticity in the cortex of the rat. Neurosci. 80: 1159–1169.
tional Physiological Congress, 88–89. Copenhagen: Bianco
Baudry, M. (1998). Synaptic plasticity and learning and
Lunos Bogtrykkeri.
memory: Fifteen years of progress. Neurobiol. Learn. Mem.
Barlow, H. B. (1972). Single units and sensation: A neuron 70: 113–118.
doctrine for perceptual psychology? Perception 1: 371–
Baylis, G. C., Rolls, E. T., and Leonard, C. M. (1985).
394.
Selectivity between faces in the responses of a population
Barlow, H. B. (1985). The twelfth Bartlett Memorial of neurons in the cortex in the superior temporal sulcus of
Lecture: The role of single neurons in the psychology of the monkey. Brain Res. 342: 91–102.
perception. Q. J. Exp. Psychol. A37: 121–145.
Baylis, G. C., Rolls, E. T., and Leonard, C. M. (1987).
Barlow, H. B. (1989). Unsupervised learning. Neural Com- Functional subdivisions of the temporal lobe neocortex. J.
put. 1: 295–311. Neurosci. 7: 330–342.
Barlow, H. B. (1990a). A theory about the functional role Beale, J. M., and Keil, F. C. (1995). Categorical e¤ects in
and synaptic mechanisms of visual after-e¤ects. In Vision: the perception of faces. Cognition 57: 217–239.
Coding and E‰ciency, ed. C. Blakemore, 363–375. Cam-
Beard, B. L., Klein, S. A., Ahumada, Jr. A. J., and Slotnick,
bridge: Cambridge University Press.
S. D. (1996). Training on a vernier acuity task does transfer
Barlow, H. B. (1990b). Conditions for versatile learning, to untrained retinal locations. Invest. Ophthalmol. Vis. Sci.
Helmholtz’s unconscious inference, and the task of per- 37: S696, 3180.
ception. Vision Res. 30: 1561–1571.
Beard, B. L., Levi, D. M., and Reich, L. N. (1995). Per-
Barlow, H. B. (1994). What is the computational goal ceptual learning in parafoveal learning. Vision Res. 35:
of the neocortex? In C. Koch and J. L. Davis eds., Large- 1679–1690.
Scale Neuronal Theories of the Brain. Cambridge, MA: MIT
Beaulieu, C., Kisvarday, Z., Somogyi, P., Cynader, M.,
Press.
and Cowey, A. (1992). Quantitative distribution of GABA-
Barlow, H. B. (1995). The neuron doctrine in perception. immunopositive and -immunonegative neurons and syn-
In M. Gazzaniga, ed., The Cognitive Neurosciences, 415–435. apses in the monkey striate cortex (area 17). Cereb. Cortex
Cambridge, MA: MIT Press. 2: 295–309.
Barlow, H. B., and Tolhurst, D. J. (1992). Why do you Bedford, F. L. (1993). Perceptual and cognitive spatial
have edge detectors? Opt. Soc. Am. Tech. Digest 23: 172. learning. J. Exp. Psychol. Hum. Percept. Perform. 19: 517–530.
390 References
Bedford, F. L. (1995). Constraints on perceptual learning: Bennett, T. L., Levitt, L., and Anton, B. S. (1972). E¤ect
Objects and dimensions. Cognition 54: 253–297. of exposure to a single stimulus on transfer of perceptual
learning. Percept. Mot. Skills 34: 559–562.
Bedford, F. L. (1997). Are long-term changes to percep-
tion explained by Pavlovian associations or perceptual Benson, P. J., and Perrett, D. I. (1991a). Synthesising
learning theory? Cognition 64: 223–230. continuous-tone caricatures. Image Vis. Comput. 9: 123–
129.
Bell, A. J. (1992). Self-organisation in real neurons: Anti-
Hebb in ‘‘channel space’’? In J. Moody et al., eds., Advances Benson, P. J., and Perrett, D. I. (1991b). Perception and
in Neural Information Processing Systems, vol. 4, pp. 59–66. recognition of photographic quality facial caricatures: Im-
San Francisco: Morgan-Kaufmann. plications for the recognition of natural images. Eur. J.
Cogn. Psychol. 3: 105–135.
Bell, A. J., and Sejnowski, T. J. (1995a). An information
maximization approach to blind separation and blind Berardi, N., and Fiorentini, A. (1987). Interhemispheric
deconvolution. Neural Comput. 7: 1129–1159. transfer of visual information in humans: Spatial charac-
teristics. J. Physiol. (Lond.) 384: 633–647.
Bell, A. J., and Sejnowski, T. J. (1995b). Fast blind sepa-
ration based on information theory. In Proceedings of the Berardi, N., and Fiorentini, A. (1991). Visual field asym-
International Symposium on Nonlinear Theory and Applications, metry in pattern discrimination: A sign of asymmetry in cor-
Las Vegas, December 1995. tical visual field representation? Vision Res. 31: 1831–1836.
Bell, A. J., and Sejnowski, T. J. (1996). Learning the Berardi, N., and Fiorentini, A. (1997). Interhemispheric
higher-order structure of a natural sound. Network 7: 2. transfer of spatial and temporal frequency information. In:
Cerebral Asymmetries in Sensory and Perceptual Processing, ed.
Bell, A. J., and Sejnowski, T. J. (1997). The ‘‘independent
S. Christman. Amsterdam: Elsevier.
component’’ of natural scenes are edge filters. Vision Res.
37: 3327–3338. Berkeley, G. (1713). Three Dialogues Between Hylas and
Philonous. London: Innys. Reproduced in A. C. Fraser, ed.,
Bellman, R. E. (1961). Adaptive Control Processes. Prince-
The Works of George Berkeley (1901). Oxford: Oxford Uni-
ton, NJ: Princeton University Press.
versity Press.
Benardete, E. A., and Kaplan, E. (1999). The dynamics of
Berkeley, G. (1732). Alciphron: or, the Minute Philosopher.
primate M retinal ganglion cells. Vis. Neurosci. 16: 355–368.
London: J. Tonson. Reproduced in A. C. Fraser, ed., The
Bende, M., and Nordin, S. (1997). Perceptual learning in Works of George Berkeley (1901). Oxford: Oxford University
olfaction: Professional wine tasters versus controls. Physiol. Press.
Behav. 62: 1065–1070.
Berlucchi, G., and Rizzolatti, G. (1968). Binocularly driven
Benedetti, F. (1991). Perceptual learning following a long- neurons in visual cortex of split-chiasm cats. Science 159:
lasting tactile reversal. J. Exp. Psychol. Hum. Percept. Perform. 308–310.
17: 267–277.
Berman, N. E., and Payne, B. R. (1983). Alterations in
Bennett, C. H., and Mackintosh, N. J. (1999). Comparison connections of the corpus callosum following convergent
and contrast as a mechanism of perceptual learning? Q. J. and divergent strabismus. Brain Res. 274: 201–212.
Exp. Psychol. B. 52: 253–272.
Berns, G. S., Cohen, J. D., and Mintun, M. A. (1997). Brain
Bennett, P. J., and Banks, M. S. (1987). Sensitivity loss in regions responsive to novelty in the absence of awareness.
odd-symmetric mechanisms underlies phase anomalies in Science 276: 1272–1276.
peripheral vision. Nature 326: 873–876.
Berry, D. C. (1994). Implicit learning: Twenty-five years
Bennett, T. L., and Anton, B. S. (1972). Critical periods on a tutorial. In C. Umiltà and M. Moscovitch, eds., Atten-
for early experience in transfer of perceptual learning. tion and Performance XV, 755–781. Cambridge, MA: MIT
Percept. Mot. Skills 35: 743–746. Press.
References 391
Bertini, G., Karni, A., De Weerd, P., Desimone, R., and Bjordahl, T. S., Dimyan, M. A., and Weinberger, N. M.
Ungerleider, L. (1995). A behavioral and electrophysi- (1998). Induction of long-term receptive field plasticity in
ological study of monkey visual cortex plasticity. Soc. the auditory cortex of the waking guinea pig by stimula-
Neurosc. Abstr. 21: 276. tion of the nucleus basalis. Behav. Neurosci. 112: 467–479.
Bertini, G., Karni, A., De Weerd, P., Desimone, R., and Blakemore, C., and Campbell, F. W. (1969). On the ex-
Ungerleider, L. (1996). Electro-physiological study of istence of neurones in the human visual system selectively
monkey visual cortex plasticity: Comparison of V1 and sensitive to the orientation and size of retinal images. J.
V2. Soc. Neurosci. Abstr. 22: 1614. Physiol. (Lond.) 203: 237–260.
Best, F. (1900). Ueber die Grenze der Erkennbarkeit Blakemore, C., and Cooper, G. F. (1970). Development of
von Lagenunterschieden. Graefes Arch. Ophthal. 51: 453– the brain depends on the visual environment. Nature 228:
460. 477–478.
Beymer, D., and Poggio, T. A. (1996). Image representa- Blakemore, C., Diao, Y., Pu, M., Wang, Y., and Xiao, Y.
tions for visual learning. Science 272: 1905–1909. (1983). Possible functions of interhemispheric connections
between visual cortical areas in the cat. J. Physiol. (Lond.)
Bialek, W., Ruderman, D. L., and Zee, A. (1991). Opti-
337: 331–348.
mal sampling of natural images: A design principle for the
visual system? In D. Touretzky, ed., Advances in Neural In- Blakemore, C. B., and Nachmias, J. (1971). The orien-
formation Processing Systems, vol. 1. San Francisco: Morgan- tational specificity of two visual after-e¤ects. J. Physiol.
Kaufmann. (Lond.) 213: 157–174.
Biederman, I. (1987). Recognition-by-components: A Blakemore, C. B., Nachmias, J., and Sutton, P. (1970).
theory of human image understanding. Psychol. Rev. 94: The perceived spatial frequency shift: Evidence for fre-
115–147. quency-selective neurons in the human brain. J. Physiol.
(Lond.) 210: 727–750.
Biederman, I., and Gerhardstein, P. C. (1993). Recogniz-
ing depth-rotated objects: Evidence and conditions for Blakemore, C. B., and Sutton, P. (1969). Size adaptation:
three-dimensional viewpoint invariance. J. Exp. Psychol. A new after-e¤ect. Science 166: 245–247.
Hum. Percep. Perform. 19: 1162–1182.
Blakemore, C., and Tobin, E. A. (1972). Lateral inhibition
Biedermann, I., and Shi¤rar, M. M. (1995). Sexing day- between orientation detectors in the cat’s visual cortex.
old chicks: A case study and expert systems analysis of a Exp. Brain Res. 15: 439–440.
di‰cult perceptual-learning task. J. Exp. Psychol. Learn.
Blamey, P., Arndt, P., Bergeron, S., Bredberg, G., Brima-
Mem. Cogn. 13: 640–645.
combe, J., Facer, G., Larky, J., Lindström, B., Nedzelski,
Bienenstock, E. L., Cooper, L. N., and Munro, P. W. J., Peterson, A., Shipp, D., Staller, S., and Whitford, L.
(1982). Theory for the development of neuron selectivity: (1996). Factors a¤ecting auditory performance of post-
Orientation specificity and binocular interaction in visual linguistically deaf adults using cochlear implants. Audiol.
cortex. J. Neurosci. 2: 32–48. Neuro-otol. 1: 293–306.
Birbaumer, N., Lutzenberger, W., Monotya, P., Larbig, Blamey, P. J., Pyman, B. C., Gordon, M., Clark, G. M.,
W., Unertl, K., Töpfner, S., Grodd, W., Taub, E., and Brown, A. M., Dowell, R. C., and Hollow, R. D. (1992).
Flor, H. (1997). E¤ects of regional anesthesia on phantom Factors predicting postoperative sentence scores in post-
limb pain are mirrored in changes in cortical reorganiza- linguistically deaf adult cochlear implant patients. Ann.
tion. J. Neurosci. 17(14): 5503–5508. Otol. Rhinol. Laryngol. 101: 342–348.
Bishop, C. (1995). Neural Networks for Pattern Recognition. Blasdel, G. G., and Salama, G. (1986). Voltage-sensitive
Oxford: Oxford University Press. dyes reveal a modular organization in monkey striate cor-
tex. Nature 321: 579–585.
392 References
Bolz, J., and Gilbert, C. D. (1990). The role of horizontal Bradlow, A. R, Pisoni, D. B, Akahane-Yamada, R., and
connections in generating long receptive fields in the cat Tohkura, Y. (1997). Training Japanese listeners to identify
visual cortex. Eur. J. Neurosci. 1: 263–268. English /r/ and /l/: IV. Some e¤ects of perceptual learning
on speech production. J. Acoust. Soc. Am. 101: 2299–2310.
Bonneh, Y., and Sagi, D. (1998). E¤ects of spatial config-
uration on contrast detection. Vision Res. 38: 3541–3553. Bradshaw, M., and Rogers, B. (1996). The interaction of
binocular disparity and motion parallax in the computation
Bonneh, Y., and Sagi, D. (1999). Configuration saliency
of depth. Vision Res. 36(21): 3457–3468.
revealed in short duration binocular rivalry. Vision Res. 39:
271–281. Brady, M., and Yuille, A. (1983). An extremum principle
for shape from contour. AI Memo 711. MIT AI Laboratory.
Booth, M. C. A., and Rolls, E. T. (1998). View-invariant
representations of familiar objects by neurons in the infe- Brady, M., and Yuille, A. (1984). An extremum principle
rior temporal visual cortex. Cereb. Cortex 8: 510–523. for shape from contour. IEEE Trans. Patt. Anal. Mach.
Intell. 6: 288–301.
Boring, E. G. (1930). A new ambiguous figure. Am. J.
Psychol. 42: 444–445. Braitenberg, V. (1986). Two views of the cerebral cortex.
In: G. Palm, and A. Aertsen eds., Brain Theory, 81–96.
Bornstein, M. H., and Korda, N. O. (1984). Discrimina-
Berlin: Springer.
tion and matching within and between hues measured by
reaction times: Some implications for categorical percep- Brasil-Neto, J. P., Cohen, L. G., Pascual-Leone, A., Jabir,
tion and levels of information processing. Psychol. Res. 46: F. K., Wall, R. T., and Hallett, M. (1992). Rapid revers-
207–222. ible modulation of human motor outputs after transient
dea¤erentation of the forearm. Neurology 42: 1302–1306.
Bosking, W. H., Zhang, Y., Schofield, B., and Fitzpatrick,
D. (1997). Orientation selectivity and the arrangement Brasil-Neto, J. P., Valls-Solé, J., Pascual-Leone, A., Cam-
of horizontal connections in tree shrew striate cortex. J. marota, A., Amassian, V. E., Cracco, R., Maccabee, P.,
Neurosci. 17: 2112–2127. Cracco, J., Hallett, M., and Cohen, L. G. (1993). Rapid
modulation of human cortical motor outputs following
Boutet, I., Intriligator, J., and Rivest, J. (1995). The in-
ischemic nerve block. Brain 116: 511–525.
fluence of attention on visual learning. Invest. Ophthalmol.
Vis. Sci. 35: 1769. Braun, C., Schweizer, R., Elbert, T., Birbaumer, N., and
Taub, E. (1999). Di¤erential reorganization in somato-
Boyd, J., and Matsubara, J. (1991). Intrinsic connections in
sensory cortex for di¤erent discrimination tasks. J. Neurosci.
cat visual cortex: A combined anterograde and retrograde
20: 446–450.
tracing study. Brain Res. 560: 207–215.
Braun, J., and Sagi, D. (1991). Texture-based tasks are little
Braastad, B. O., and Heggelund, P. (1985). Development
a¤ected by second tasks requiring peripheral or central at-
of spatial receptive-field organization and orientation se-
tentive fixation. Perception 20: 483–500.
lectivity in kitten striate cortex. J. Neurophysiol. 53: 1158–
1178. Bregler, C., and Omohundro, S. M. (1995). Nonlinear
image interpolation using manifold learning. In D. S. T.
Braddick, O., Campbell, F. W., and Atkinson, J. (1978).
Tesauro and T. K. Leen, eds., Advances in Neural Informa-
Channels in vision. In Handbook of Sensory Physiology. Vol.
tion Processing, vol. 7, pp. 973–980. Cambridge, MA: MIT
8, Perception, ed. R. Held, H. W. Leibowitz, and H. L.
Press.
Teuber, 1–38 Berlin: Springer.
Brennan, S. E. (1985). The caricature generator. Leonardo
Bradley, A., Switkes, E., and De Valois, K. (1988).
18: 170–178.
Orientation and spatial frequency selectivity of adapta-
tion to colour and luminance gratings. Vision Res. 28: Brigham, J. C. (1986). The influence of race on face
841–859. recognition. In H. D. Ellis, M. A. Jeeves, F. Newcombe,
References 393
and A. Young, eds., Aspects of Face Processing. Dordrecht: Bruce, V., Carson, D., Burton, A. M., and Ellis, A. W.
Nijho¤. (2000). Perceptual priming is not a necessary consequence
of semantic classification of pictures. Q. J. Exp. Psychol. A.
Broadbent, D. E. (1956). Successive responses to simulta-
53: 289–323.
neous stimuli. Q. J. Exp. Psychol. 8: 145–152.
Bruce, V., Carson, D., Burton, A. M., and Kelly, S. (1998).
Brodlie, J. F., and Burke, J. (1971). Perceptual learning
Prime-time advertisements: Repetition priming of faces
disabilities in blind children. Percept. Mot. Skills 32: 313–
seen on recruitment posters Mem. Cognit. 26: 502–515.
314.
Bruce, V., Doyle, T., Dench, N., and Burton, A. M.
Brodmann, K. (1909). Vergleichende Lokalisationslehre der
(1991). Remembering facial configurations. Cognition 38:
Großhirnrinde in ihren Prinzipien dargestellt auf Grund des
109–144.
Zellenbaues. Leipzig: Barth.
Bruce, V., Hanna, E., Dench, N., Healey, P., and Burton,
Brooks, R. (1981). Symbolic reasoning among 3-D and
M. (1992). The importance of ‘‘mass’’ in line-drawings of
2-D images. Artif. Intell. 17: 205–244.
faces. Appl. Cogn. Psychol. 6: 619–628.
Brosvic, G. M, Rowe-Boyer, M. M., and Diho¤, R. E.
Bruce, V., Healey, P., Burton, A. M., Doyle, T., Coombes,
(1991). Acquisition and retention of perceptual learning
A., and Linney, A. (1991). Recognising facial surfaces.
and the horizontal-vertical illusion. Percept. Mot. Skills 72:
Perception 20: 755–769.
587–592.
Bruce, V., Henderson, Z., Greenwood, K., Hancock,
Brown, T. H., Kairiss, E. W., and Keenan, C. L. (1990).
P. J. B., Burton, A. M., and Miller, P. (1999). Verification of
Hebbian synapses: Biophysical mechanisms and algorithms.
face identities from images captured on video. J. Exp.
Annu. Rev. Neurosci. 13: 475–511.
Psychol. A.
Bruce, C., Desimone, R., and Gross, C. G. (1981). Visual
Bruce, V., Henderson, Z., Newman, C., and Burton, A.
properties of neurons in a polysensory area in superior
M. (2001). Matching identities of familiar and unfamiliar
temporal sulcus of the macaque. J. Neurophysiol. 46: 369–
faces caught on CCTV images. J. Exp. Psychol. A. 7: 207–
384.
218.
Bruce, V. (1982). Changing faces: Visual and non-visual
Bruce, V., and Humphreys, G. W. (1994). Recognising
coding processes in face recognition. Br. J. Psychol. 73:
objects and faces. Vis. Cognit. 1: 141–180.
105–116.
Bruce, V., and Langton, S. (1994). The use of pigmen-
Bruce, V. (1994). Stability from variation: The M. D.
tation and shading information in recognising the sex and
Vernon memorial lecture. Q. J. Exp. Psychol. 47A: 5–28.
identities of faces. Perception 23: 803–822.
Bruce, V. (1998). Fleeting images of shade: Identifying
Bruce, V., Terry, D., and Smith, K. (1998). Visual and
people caught on video. Psychologist 11: 331–337.
non-visual factors a¤ecting repetition priming of faces.
Bruce, V., Burton, A. M., Carson, D., Hanna, E., and Paper presented at the Tenth Congress of the European
Mason, O. (1994). Repetition priming of face recognition. Society for Cognitive Psychology, Jerusalem, September.
In Carlo Umiltà and Morris Moskovitch, eds., Attention
Bruce, V., and Young, A. W. (1986). Understanding face
and Performance XV. Cambridge, MA: MIT Press.
recognition. Br. J. Psychol. 77: 305–327.
Bruce, V., Campbell, R. N., Doherty-Sneddon, G., Im-
Bruner, J. S. (1951). Personality dynamics and the process
port, A., Langton, S., McAuley, S., and Wright, R. (2000).
of perceiving. In R. R. Blake and G. V. Ramsey, eds.,
Testing face processing skills in children. Br. J. Dev. Psy-
Perception: An Approach to Personality. New York: Ronald
chol. 18: 319–333.
Press.
394 References
Bruner, J. S. (1957). On perceptual readiness. Psychol. Rev. not sensitive to unpaired presynaptic activity. J. Neuro-
64: 123–152. physiol. 76: 631–636.
Bruner, J. S., and Goodman, C. C. (1947). Value and need Buonomano, D. V., and Merzenich, M. M. (1998a). Cor-
as organizing factors in perception. J. Abnorm. Soc. Psychol. tical plasticity: From synapses to maps. Annu. Rev. Neurosci.
42: 33–44. 21: 149–186.
Bruner, J. S., and Postman, L. (1947). Emotional selectiv- Buonomano, D. V., and Merzenich, M. M. (1998b). Net
ity in perception and reaction. J. Pers. 16: 69–77. interaction between di¤erent forms of short-term synaptic
plasticity and slow IPSPs in the hippocampus and auditory
Büchel, C., Price, C., Frackowiak, R. S. J., and Friston, K.
cortex. J. Neurophysiol. 80: 1765–1774.
(1998). Di¤erent activation patterns in the visual cortex of
late and congenitally blind subjects. Brain 121: 404–419. Burkhalter, A., Bernardo, K. L., and Charles, V. (1993).
Development of local circuits in human visual cortex. J.
Buchner, H., Reinartz, U., Waberski, T. D., Gobbele, R.,
Neurosci. 13: 1916–1931.
Noppeney, U., and Scherg, M. (1999). Sustained attention
modulates the immediate e¤ect of de-a¤erentiation on the Burr, D. C. (1980). Sensitivity to spatial phase. Vision Res.
cortical representation of the digits: Source localization of 20: 391–396.
somatosensory evoked potentials in humans. Neurosci. Lett.
Burr, D. C., Morrone, M. C., and Spinelli, D. (1989).
260: 57–60.
Evidence for edge and bar detectors in human vision. Vision
Buhmann, J., Lades, M., and von der Malsburg, C. (1990). Res. 29: 419–431.
Size- and distortion-invariant object recognition by hier-
Burr, D. C., and Wijesundra, S. (1991). Orientation dis-
archical graph matching. In International Joint Conference on
crimination depends on spatial frequency. Vision Res. 31:
Neural Networks, 411–416. New York: IEEE Press.
1449–1452.
Bültho¤, H. H., and Edelman, S. (1992). Psychophysical
Burton, A. M. (1994). Learning new faces in an interactive
support for a two-dimensional view interpolation theory
activation and competition model. Vis. Cognit. 1: 313–
of object recognition. Proc. Natl. Acad. Sci. U S A 89: 60–
348.
64.
Burton, A. M. (1998). A model of human face recognition.
Bültho¤, H., and Yuille, A. (1996). A Bayesian framework
In J. Grainger and A. M. Jacobs, eds., Localist Connectionist
for the integration of visual modules. In J. McClelland and
Approaches to Human Cognition, 75–100. Mahwah, NJ:
T. Inui, eds., Attention and Performance XVI: Information
Erlbaum.
Integration in Perception and Communication, 49–70. Cam-
bridge, MA: MIT Press. Burton, A. M., Bruce, V., and Hancock, P. J. B. (1999).
From pixels to people: A model of familiar face recogni-
Bültho¤, I., Bültho¤, H. H., and Sinha, P. (1998). Top-
tion. Cogn. Sci. 23: 1–31.
down influences on stereoscopic depth-perception. Nat.
Neurosci. 1: 254–257. Burton, A. M., Bruce, V., and Johnston, R. A. (1990).
Understanding face recognition with an interactive acti-
Bültho¤, I., Sinha, P., and Bültho¤, H. H. (1996). Top-
vation model. Br. J. Psychol. 81: 361–380.
down influence of recognition on stereoscopic depth per-
ception. Invest. Ophthalmol. Vis. Sci. 37: 5168. Burton, A. M., Wilson, S., Cowan, M., and Bruce, V.
(1999). Face recognition in poor-quality video: Evidence
Buonomano, D. V. (1999). Distinct functional types of
from security surveillance Psychol. Sci. 10: 243–248.
associative long-term potentiation in neocortical and hip-
pocampal pyramidal neurons. J. Neurosci. 19: 6748–6754. Busby, P. A., and Clark, G. M. (2000). Electrode dis-
crimination by early deafened subjects using the Cochlear
Buonomano, D. V., and Merzenich, M. M. (1996). Asso-
Limited multiple-electrode cochlear implant. Ear Hear. 21:
ciative synaptic plasticity in hippocampal CA1 neurons is
291–304.
References 395
Busby, P. A., and Clark, G. M. (2000). Pitch estimation by Campbell, F. W., and Kulikowski, J. J. (1966). Orientation
early-deafened subjects using a multiple-electrode cochlear selectivity of the human visual system. J. Physiol. 187: 437–
implant. J. Acoust. Soc. Am. 107: 547–558. 445.
Bushnell, I. W. R., Sai, F., and Mullin, J. T. (1989). Campbell, F. W., and Ma¤ei, L. (1971). The tilt after-
Neonatal recognition of the mother’s face. Br. J. Dev. e¤ect: A fresh look. Vision Res. 11: 833–840.
Psychol. 7: 3–15.
Campbell, F. W., Nachmias, J., and Hukes, J. (1970).
Byl, N. N., Merzenich, M. M., Cheung, S., Bedenbaugh, Spatial frequency discrimination in human vision. J. Opt.
P., Nagarajan, S. S., and Jenkins, W. M. (1997). A primate Soc. Am. 60: 555–559.
model for studying focal dystonia and repetitive strain in-
Candia, V., Elbert, T., Altenmüller, E., Rau, H., Schäfer,
jury: E¤ects on the primary somatosensory cortex. Phys.
T., and Taub, E. (1998). A constraint-induced movement
Ther. 77: 269–284.
therapy for focal hand dystonia in musicians. Lancet 353:
Byl, N. N., Merzenich, M. M., and Jenkins, W. M. (1996). 52.
A primate genesis model of focal dystonia and repetitive
Canny, J. F. (1986). A computational approach to edge-
strain injury: 1. Learning-induced dedi¤erentiation of the
detection. IEEE Trans. Patt. Anal. Mach. Vis. 8: 679–698.
representation of the hand in the primary somatosensory
cortex in adult monkeys. Neurology 47: 508–520. Carandini, M., and Ferster, D. (1997). A tonic hyper-
polarisation underlying contrast adaptation in cat visual
Cabeza, R., Bruce, V., Kato, T., and Oda, M. (1999). The
cortex. Science 276: 949–952.
prototype e¤ect in face recognition: Extension and limits.
Mem. Cognit. 27: 139–151. Carandini, M., Heeger, D. J., and Movshon, J. A. (1997).
Linearity and normalization in simple cells of the macaque
Caelli, T. (1997). Perceptual learning and adaptation in
primary visual cortex. J. Neurosci. 17: 8621–8644.
man and machine: Part II. Spatial Vision 10: 505–508.
Carbonell, J. G., Michalski, R. S., and Mitchell, T. M.
Cahusac, P. M. (1995). Synaptic plasticity induced in single
(1983). An overview of machine learning. In R. S.
neurones of the primary somatosensory cortex in vivo. Exp.
Michalski, J. G. Carbonell, and T. M. Mitchell, eds., Ma-
Brain Res. 107: 241–253.
chine Learning: An Artificial Intelligence Approach, 3–23. Palo
Calford, M. B., and Tweedale, R. (1988). Immediate Alto, CA: Tioga.
chronic changes in responses of somatosensory cortex in
Cardoso, J.-F., and Laheld, B. (1996). Equivariant adaptive
adult flying-fox after digit amputation. Nature 332: 446–
source separation. IEEE Trans. Signal Proc. 45: 434–444.
448.
Carew, T. J., Hawkins, R. D., Abrams, T. W., and Kandel,
Calford, M. B., and Tweedale, R. (1990). Interhemi-
E. R. (1984). A test of Hebb’s postulate at identified syn-
spheric transfer of plasticity in the cerebral cortex. Science
apses which mediate classical conditioning in Aplysia. J.
249: 805–807.
Neurosci. 4: 1217–1224.
Calford, M. B., and Tweedale, R. (1991). Immediate ex-
Carey, S., and Diamond, R. (1994). Are faces perceived as
pansion of receptive fields of neurons in area 3b of ma-
configurations more by adults than by children? Vis. Cognit.
caque monkeys after digit denervation. Somatosens. Mot.
1: 253–274.
Res. 8(3): 249–260.
Carpenter, G. A., and Grossberg, S. (1987). A massively
Callaway, E. M., and Katz, L. C. (1990). Emergence and
parallel architecture for a self-organizing neural pattern
refinement of clustered horizontal connections in cat
recognition machine. Comput. Vis. Graph. Image Proc. 37:
striate cortex. J. Neurosci. 10: 1134–1153.
54–115.
Campbell, F. W., and Gubisch, R. W. (1966). Optical
Carpenter, G. A., and Grossberg, S. (1990). Adaptive
quality of the human eye. J. Physiol. 186: 558–578.
resonance theory: Neural network architectures for self-
396 References
organizing pattern recognition. In R. Eckmiller, G. Christie, F., and Bruce, V. (1998). The role of movement
Hartmann, and G. Hauske, eds., Parallel Processing in Neu- in the recognition of unfamiliar faces. Mem. Cognit. 26:
ral Systems and Computers, 383–389. Amsterdam: North- 780–790.
Holland.
Chung, M. S., and Thomson, D. M. (1995). Development
Carpenter, G. A., Grossberg, S., Markuzon, N., Reynolds, of face recognition. Br. J. Psychol. 86: 55–87.
J. H., and Rosen, D. B. (1992). Fuzzy ARTMAP: A neural
Churchill, J. P., Muja, N., Myers, W., Besheer, W. A., and
network architecture for incremental supervised learning
Garraghty, P. E. (1998). Somatotopic consolidation: A third
of analog multidimensional maps. IEEE Trans. Neural Net-
phase of cortical reorganization after peripheral nerve injury
works, 3: 698–713.
in adult squirrel monkeys. Exp. Brain Res. 118: 189–194.
Cavanagh, P. (1991). In Representations of Vision: Trends
Churchland, P. S., and Sejnowski, T. J. (1988). Perspectives
and Tacit Assumptions in Vision Research, ed. A. Gorea.
on cognitive neuroscience. Science 242: 741–745.
Cambridge: Cambridge University Press.
Churs, L., Spengler, F., Jürgens, M., and Dinse, H. R.
Changeux, J.-P., and Danchin, A. (1976). Selective stabi-
(1996). Environmental enrichment counteracts decline of
lisation of developing synapses as a mechanism for the
sensorimotor performance and deterioration of cortical
specification of neuronal networks. Nature 264: 705–712.
organization in aged rats. Soc. Neurosci. Abstr. 22: 102.
Chen, R., Corwell, B., Yaseen, Z., Hallett, M., and Cohen,
Cichocki, A., Unbehauen, R., and Rummert, E. (1994).
L. G. (1998). Mechanisms of cortical reorganization in
Robust learning algorithm for blind separation of signals.
lower-limb amputees. J. Neurosci. 18(9): 3443–3450.
Electron. Lett. 30: 1386–1387.
Cheng, K., Saleem, K. S., and Tanaka, K. (1997). Orga-
Clark, A. (1993). Sensory Qualities. Oxford: Clarendon Press.
nization of corticostriatal and corticoamygdalar projections
arising from the anterior inferotemporal area TE of the Clark, G. M. (1986). The University of Melbourne/
macaque monkey: A Phaseolus vulgaris leucoagglutinin Cochlear Corporation (Nucleus) program. Otolaryngol.
study. J. Neurosci. 15: 7902–7925. Clin. North Am. 19: 329–354.
Chino, Y. M. (1995). Adult plasticity in the visual system. Clark, G. M. (1996). Electrical stimulation of the auditory
Can. J. Physiol. Pharmacol. 73: 1323–1338. nerve: The coding of frequency, the perception of pitch,
and the development of cochlear implant speech process-
Chino, Y. M. (1997). Receptive-field plasticity in the
ing strategies for profoundly deaf people. J. Clin. Physiol.
adult visual cortex: Dynamic signal rerouting or experi-
Pharm. Res. 23: 766–776.
ence-dependent plasticity. Semin. Neurosci. 9: 34–46.
Clark, G. M. (1997). Auditory nervous system plasticity:
Chino, Y. M., Kaas, J. H., Smith III, E. L., Langston, A.
Application to cochlear implantation. In Abstract Book for
L., and Cheng, H. (1992). Rapid reorganization of cortical
Sixteenth World Congress of Otorhinolaryngology Head and
maps in adult cats following restricted dea¤erentation in
Neck Surgery, Sydney, March 2–7, 1997, vol. 20, pp. 19–23.
retina. Vision Res. 32: 789–796.
Clark, S. A., Allard, T., Jenkins, W. M., and Merzenich,
Chiroro, P., and Valentine, T. (1995). An investigation of
M. M. (1988). Receptive fields in the body-surface map in
the contact hypothesis of the own-race bias in face recog-
adult cortex defined by temporally correlated inputs. Nature
nition. Q. J. Exp. Psychol. 48A: 879–894
332: 444–445.
Choi, D. W., and Rothman, S. M. (1990). The role of
Clothiaux, E. E., Bear, M. F., and Cooper, L. N. (1991).
glutamate neurotoxicity in hypoxic-ischemic neuronal
Synaptic plasticity in visual cortex: Comparison of theory
death. Annu. Rev. Neurosci. 13: 171–182.
with experiment. J. Neurophysiol. 66: 1785–1804.
Cholewiak, R. W. (1976). Satiation in cutaneous saltation.
Clowes, M. B. (1971). On seeing things. Artif. Intell. 2(1):
Sens. Processes 1: 163–175.
79–116.
References 397
Cohen, L. G., Bandinelli, S., Findley, T. W., and Hallet, Cowan, eds., International Cochlear Implant, Speech and
M. (1991a). Motor Reorganization after upper limb am- Hearing Symposium, Melbourne, 1994. St. Louis: Annals.
putation in man: A study with focal magnetic stimulation. Ann. Otol. Rhinol. Laryngol. 104, suppl. 166: 318–321.
Brain 114: 615–627.
Cowey, A. (1992). The role of the face-cell area in the
Cohen, L. G., Bandinelli, S., Sato, S., Kufta, C., and Hal- discrimination and recognition of faces by monkeys. Philos.
lett, M. (1991). Attenuation in detection of somatosensory Trans. R. Soc. London B Biol. Sci. 335: 31–38.
stimuli by transcranial magnetic stimulation. Electroenc.
Creutzfeldt, O. D., Garey, L. J., Kuroda, R., and Wol¤,
Clin. Neurophysiol. 81: 366–376.
J.-R. (1977). The distribution of degenerating axons after
Cohen, L. G., Celnik, P., Pascual-Leone, A., Corwell, B., small lesions in the intact and isolated visual cortex of the
Falz, L., Dambrosia, J., Honda, M., Sadato, N., Gerlo¤, cat. Exp. Brain Res. 27: 419–440.
C., Catala, M. D., and Hallett, M. (1997). Functional rel-
Crick, F., and Koch, C. (1998). Consciousness and neuro-
evance of cross-modal plasticity in blind humans. Nature
science. Cereb. Cortex 8: 97–107.
389: 180–183.
Crist, R. E., Kapadia, M. K., Westheimer, G., and Gilbert,
Cohen, L. G., Weeks, R. A., Sadato, N., Celnik, P., Ishii,
C. D. (1997). Perceptual learning of spatial localization:
K., and Hallett, M. (1999). Period of susceptibility for
Specificity for orientation, position, and context. J. Neuro-
cross-modal plasticity in the blind. Ann. Neurol. 45: 451–
physiol. 78: 2889–2894.
460.
Crist, R. E., Li, W., and Gilbert, C. D. (2001). Learning to
Comon, P. (1994). Independent component analysis, a
see: Experience and attention in primary visual cortex.
new concept? Signal Proc. 36: 287–314.
Nat. Neurosci. 4: 519–525.
Constantine-Paton, M., Cline, H. T., and Debski, E.
Crook, J. M., and Eysel, U. T. (1992). GABA-induced
(1990). Patterned activity, synaptic convergence, and the
inactivation of functionally characterized sites in cat visual
NMDA receptor in developing visual pathways. Annu.
cortex (area 18): E¤ects on orientation tuning. J. Neurosci.
Rev. Neurosci. 13: 129–154.
12: 1816–1825.
Coq, J. O., and Xerri, C. (1998). Environmental enrich-
Crook, J. M., Eysel, U. T., and Machemer, H. F. (1991).
ment alters organizational features of the forepaw repre-
Influence of GABA-induced remote inactivation on the
sentation in the primary somatosensory cortex of adult rats.
orientation tuning of cells in area 18 of feline visual cortex:
Exp. Brain Res. 121: 191–204.
A comparison with area 17. Neuroscience 40: 1–12.
Cottrell, G. W., Munro, P., and Zipser, D. (1987). Learn-
Crook, J. M., Kisvárday, Z. F., and Eysel, U. T. (1996).
ing internal representations from gray-scale images: An ex-
GABA-induced inactivation of functionally characterized
ample of extensional programming. In Proceedings of Ninth
sites in cat visual cortex (area 18): Local determinants of
Annual Conference of the Cognitive Science Society, 462–473.
direction selectivity. J. Neurophysiol. 75: 2071–2088.
Hillsdale, NJ: Erlbaum.
Crook, J. M., Kisvárday, Z. F., and Eysel, U. T. (1997).
Cover, T. M., and Thomas, J. A. (1991). Elements of Infor-
GABA-induced inactivation of functionally characterized
mation Theory. New York: John Wiley.
sites in cat striate cortex: E¤ects on orientation tuning and
Cowan, N., and Wood, N. L. (1997). Constraints on direction selectivity. Vis. Neurosci. 14: 141–158.
awareness, attention, processing, and memory: some recent
Crook, J. M., Kisvárday, Z. F., and Eysel, U. T. (1998).
investigations with ignored speech. Conscious. Cognit. 6:
Evidence for a contribution of lateral inhibition to orien-
182–203.
tation tuning and direction selectivity in cat visual cortex:
Cowan, R. S. C., Brown, C., Whitford, L. A., et al. (1995). Reversible inactivation of functionally characterized sites
Speech perception in children using the advanced Speak combined with neuroanatomical tracing techniques. Eur.
speech-processing system. In G. M. Clark and R. S. C. J. Neurosci. 10: 2056–2075.
398 References
Cross, K. D. (1967). Role of practice in perceptual-motor revealed by optical recording of cat primary visual cortex.
learning. Am. J. Phys. Med. 46: 487–510. Nature 375: 780–784.
Crovitz, H. F., Harvey, M. T., and McClanahan, S. (1981). Das, A., and Gilbert, C. D. (1995b). Receptive field ex-
Hidden memory: A rapid method for the study of amnesia pansion in adult visual cortex is linked to dynamic changes
using perceptual learning. Cortex 17: 273–278. in strength of cortical connections. J. Neurophysiol. 74:
779–792.
Crowley, J. C., Bosking, W. H., Foster, M., and Fitzpa-
trick, D. (1996). Development of horizontal connections Das, A., and Gilbert, C. D. (1999). Topography of con-
in layer 2/3 of tree shrew striate cortex: Relation to maps textual modulations mediated by short-range interactions
of orientation preference. Soc. Neurosci. Abstr. 22: 404.10. in primary visual cortex. Nature 399: 655–661.
Cruikshank, S. J., and Weinberger, N. M. (1996a). Evi- Daugman, J. G. (1985). Uncertainty relation for resolution
dence for the Hebbian hypothesis in experience-dependent in space, spatial frequency, and orientation optimized by
physiological plasticity of neocortex: A critical review. two-dimensional visual cortical filters. J. Opt. Soc. Am. A
Brain Res. Rev. 22: 191–228. 2(7): 1160–1169.
Cruikshank, S. J., and Weinberger, N. M. (1996b). Daum, I., and Ackermann, H. (1997). [Nondeclarative
Receptive-field plasticity in the adult auditory cortex in- memory—neuropsychological findings and neuroanatomic
duced by Hebbian covariance. J. Neurosci. 16: 861–875. principles]. Fortschr. Neurol. Psychiatr. 65: 122–132.
Cusick, C. G., Wall, J. T., Jr., Whiting, J. H., and Wiley, Davies, G., Ellis, H., and Shepherd, J. (1978). Face recog-
R. G. (1990). Temporal progression of cortical reorgani- nition accuracy as a function of mode of representation. J.
zation following nerve injury. Brain Res. 537: 355–358. Appl. Psychol. 63: 180–187.
Cutting, J. (1986). Perception with an Eye for Motion. Cam- Dawson, P. W., and Clark, G. M. (1997). Changes in
bridge, MA: MIT Press. synthetic and natural vowel perception after specific train-
ing for congenitally deafened patients using a multichannel
Cynader, M., Lepore, F., and Guillemot, J. P. (1981). Inter-
cochlear implant. Ear Hear. 18: 488–501.
hemispheric competition during postnatal development.
Nature 290: 139–140. Dayan, P., Hinton, G. E., Neal, R. M., and Zemel, R. S.
(1995). The Helmholtz machine. Neural Comput. 7: 889–
D’Amelio, F., Fox, R. A., Wu, L. C., and Daunton, N. G.
904.
(1996). Quantitative changes of GABA-immunoreactive
cells in the hindlimb representation of the rat somato- de Condillac, E. B. (1754). Treatise on the sensations. In
sensory cortex after 14-day hindlimb unloading by tail Philosophical Writings of Etienne Bonnot, abbé de Condillac.
suspension. J. Neurosci. Res. 44: 532–539. 1982. Hillsdale, NJ: Erlbaum.
Daniel, W. F, Crovitz, H. F., and Weiner, R. D. (1984). De Luca, E., and Fahle, M. (1999). Learning of interpola-
Perceptual learning with right unilateral versus bilateral tion in 2 and 3 dimensions. Vision Res. 39: 2051–2062.
electroconvulsive therapy. Br. J. Psychiatry 145: 394–400.
De Renzi, E. (1997). Prosopagnosia. In T. Feinberg and
Darian-Smith, C., and Gilbert, C. D. (1994). Axonal M. Farah, eds., Behavioral Neurology and Neuropsychology,
sprouting accompanies functional reorganization in adult 245–255. New York: McGraw-Hill.
cat striate cortex. Nature 368: 737–740.
De Valois, R. L., Yund, E. W., and Hepler, N. (1982).
Darian-Smith, C., and Gilbert, C. D. (1995). Topographic The orientation and direction selectivity of cells in ma-
reorganization in the striate cortex of the adult cat and caque visual cortex. Vision Res. 22: 531–544.
monkey is cortically mediated. J. Neurosci. 15: 1631–1647.
Dean, P. (1976). E¤ects of inferotemporal lesions on the
Das, A., and Gilbert, C. D. (1995a). Long-range horizontal behavior of monkeys. Psychol. Bull. 83: 41–71.
connections and their role in cortical reorganization
References 399
DeAngelis, G. C., Anzai, A., Ohzawa, I., and Freeman, R. Dinse, H. R. (1994). A time-based approach towards cor-
D. (1995). Receptive field structure in the visual cortex: tical functions: Neural mechanisms underlying dynamic
Does selective stimulation induce plasticity? Proc. Natl. aspects of information processing before and after post-
Acad. Sci. U S A 92: 9682–9686. ontogenetic plastic processes. Physica D75: 129–150.
Demany, L. (1985). Perceptual learning in frequency dis- Dinse, H. R., Godde, B., Hilger, T., Haupt, S. S., Spengler,
crimination. J. Acoust. Soc. Am. 78: 1118–1120. F., and Zepka, R. (1997). Short-term functional plasticity
of cortical and thalamic sensory representations and its im-
Desimone, R. (1991). Face-selective cells in the temporal
plication for information processing. Adv. Neurol. 73: 159–
cortex of monkeys. J. Cogn. Neurosci. 3: 1–8.
178.
Desimone, R. (1992). The physiology of memory:
Dinse, H. R., Recanzone, G., and Merzenich, M. M.
Recordings of things past. Science 258: 245–246.
(1990). Direct observation of neural assemblies during
Desimone, R., Albright, T. D., Gross, C. G., and Bruce, neocortical representational reorganization. In R. Eckmil-
C. (1984). Stimulus-selective properties of inferior tem- ler, G. Hartmann, and G. Hauske eds., Parallel Processing in
poral neurons in the macaque. J. Neurosci. 4: 2051–2062. Neural Systems and Computers, 65–70. Amsterdam: Elsevier.
Desimone, R., Fleming, J., and Gross, C. D. (1980). Pre- Dinse, H. R., Recanzone, G., and Merzenich, M. M.
striate a¤erents to inferior temporal cortex: An HRP (1993). Alterations in correlated activity parallel ICMS-
study. Brain Res. 184: 41–55. induced representational plasticity. Neuroreport 5: 173–
Desimone, R., and Ungerleider, L. G. (1989). Neural 176.
mechanisms of visual processing in monkey. In Handbook of Dinse, H. R., Zepka, R. F., Jürgens, M., Godde, B., Hilger,
Neuropsychology, ed. F. Boller and J. Grafman, vol. 2, pp. H., and Berkefeld, T. (1995). Age-dependent changes
267–299. Amsterdam: Elsevier. of cortical and thalamic representations revealed by opti-
Deutsch, D. (1987). The Fabric of Reality. New York: cal imaging and electrophysiological mapping techniques:
Viking Penguin. Evidence for degenerative and use-disuse-dependent pro-
cesses. In Proceedings of the C.I.N.P. Conference on Neuro-
Deutsch, J. A., and Deutsch, D. (1963). Attention: Some psychopharmacology. Homeostasis in Health and Disease
theoretical considerations. Psychol. Rev. 70: 80–90. 36(Suppl 1): 49.
Diamond, M. E., Armstrong-James, M., and Ebner, F. F. Doetsch, G. S. (1998). Perceptual significance of somato-
(1993). Experience-dependent plasticity in adult rat barrel sensory cortical reorganization following peripheral dener-
cortex. Proc. Natl. Acad. Sci. U S A 90: 2082–2086. vation. NeuroReport 9: R29–R35.
Diamond, R., and Carey, S. (1986). Why faces are and Dolan, R. J., Fink, G. R., Rolls, E., Booth, M., Holmes,
are not special: An e¤ect of expertise. J. Exp. Psychol. Gen. A., Frackowiak, R. S., and Friston, K. J. (1997). How
115: 107–117. the brain learns to see objects and faces in an impoverished
Dill, M., and Fahle, M. (1997). The role of visual field context. Nature 389: 596–599.
position in pattern-discrimination learning. Proc. R. Soc. Domann, R., Hagemann, G., Kraemer, M., Freund, H. J.,
Lond. B Biol. Sci. 264: 1031–1036. and Witte, O. W. (1993). Electrophysiological changes in
Dill, M., and Fahle, M. (1999). Display symmetry a¤ects the surrounding brain tissue of photochemically induced
positional specificity in same-di¤erent judgement of pairs cortical infarcts in the rat. Neurosci. Lett. 155: 69–72.
of novel visual patterns. Vision Res. 39: 3752–3760. Doniger, G. M., Foxe, J. J., Schroeder, C. E., Murray, M.
Dill, M., and Heisenberg, M. (1995). Visual pattern M., Higgins, B. A., and Javitt, D. C. (2001). Visual per-
memory without shape recognition. Philos. Trans. R. Soc. ceptual learning in human object recognition areas: A
Lond. B Biol. Sci. 349: 143–152. repetition priming study using high-density electrical
mapping. Neuroimage 13: 305–313.
400 References
Donoghue, J. P. (1995). Plasticity of adult sensorimotor Duke-Elder, S., and Wybar, K. (1973). System of Ophthal-
representations. Curr. Opin. Neurobiol. 5: 749–754. mology. Vol. 6, Ocular Motility and Strabismus. London:
Kimpton.
Dorais, A., and Sagi, D. (1997). Contrast masking e¤ects
change with practice. Vision Res. 37: 1725–1733. Duncan, P. (1997). Synthesis of intervention trials to im-
prove motor recovery following stroke. Top. Stroke Rehab.
Dosher, B. A., and Lu, Z. L. (1998). Perceptual learning
3(4): 1–20.
reflects external noise filtering and internal noise reduction
through channel reweighting. Proc. Natl. Acad. Sci. U S A Durack, J. C., and Katz, L. C. (1996). Development of
95: 13988–13993. horizontal projections in layer 2/3 of ferret visual cortex.
Cereb. Cortex 6: 178–183.
Dosher, B. A., and Lu, Z. L. (1999). Mechanisms of per-
ceptual learning. Vision Res. 39: 3197–3221. Duvdevani-Bar, S., Edelman, S., Howell, A. J., and
Buxton, H. (1998). A similarity-based method for the gen-
Dowell, R. C., Brown, A. M., Seligman, P. M., and Clark,
eralization of face recognition over pose and expression.
G. M. (1985). Patient results for a multiple-channel coch-
In S. Akamatsu and K. Mase, eds., Proceedings of the Third
lear prosthesis. In R. A. Schindler and M. M. Merzenich,
International Symposium on Face and Gesture Recognition,
eds., Cochlear Implants (Tenth Anniversary Conference on
118–123. Washington, DC: IEEE Press.
Cochlear Implants. San Francisco, June 22–24, 1983),
421–431. New York: Raven Press. Dykes, R. W., Landry, P., Metherate, R., and Hicks, T. P.
(1984). Functional role of GABA in cat primary somato-
Dowell, R. C., Dettman, S. J., and Barker, E. J. (1998).
sensory cortex: Shaping receptive fields of cortical neu-
Long-term outcomes for children using cochlear implants.
rons. J. Neurophysiol. 52: 1066–1093.
Paper presented at Bi-annual Scientific Meeting of the
Audiological Society of Australia, Sydney, April 1998. Eagleman, D. M., Coenen, O. J. M.-D., Mitsner, V.,
Aust. J. Audiol. Suppl. 20: 67. Bartol, T. M., Bell, A. J., and Sejnowski, T. J. (2001).
Cerebellar glomeruli: Does limited extracellular calcium
Dowell, R. C., Mecklenburg, D. J., and Clark, G. M.
implement a sparse coding strategy? In Proceedings of the
(1986). Speech recognition for 40 patients receiving mul-
8th Joint Sumposium on Neural Computation. http://www.its.
tichannel cochlear implants. Arch. Otolaryngol. 112: 1054–
caltech.edu/~jsnc/Proceedings/Eagleman-D.pdf.
1059.
Ebbinghaus, H. (1885). Memory: A Contribution to Experi-
Downing, P. E., and Treisman, A. M. (1997). The line-
mental Psychology, Trans. H. A. Ruger and C. E. Bussenius.
motion illusion: Attention or impletion? J. Exp. Psychol.
1913. Reprint, New York: Dover, 1964.
Hum. Percept. Perform. 23: 768–779.
Eddington, D. K. (1980). Speech discrimination in deaf
Drasdo, N. (1991). Neural substrates and threshold gra-
subjects with cochlear implants. J. Acoust. Soc. Am. 68:
dients of peripheral vision. In J. J. Kulikowski, V. Walsh,
885–891.
and I. J. Murray, eds., Limits of Vision. Boca Raton, FL:
Macmillan. Edeline, J. M. (1996). Does Hebbian synaptic plasticity
explain learning-induced sensory plasticity in adult mam-
Dresp, B. (1999). Dynamic characteristics of spatial mech-
mals? J. Physiol. Paris 90: 271–276.
anisms coding contour structures. Spat. Vis. 12: 129–142.
Edeline, J. M., Hars, B., Maho, C., and Hennevin, E.
Dresp, B., and Bonnet, C. (1991). Psychophysical evidence
(1994). Transient and prolonged facilitation of tone-
for low-level processing of illusory contours and surfaces in
evoked responses induced by basal forebrain stimula-
the Kanizsa square. Vision Res. 31: 1813–1817.
tions in the rat auditory cortex. Exp. Brain Res. 97: 373–
Dretske, F. (1995). Naturalizing the Mind. Jean Nicod 386.
Lectures. Cambridge, MA: MIT Press.
Edelman, G. M. (1987). Neural Darwinism: The Theory of
Duda, R. O., and Hart, P. E. (1973). Pattern Classification Neuronal Group Selection. New York: Basic Books.
and Scene Analysis. New York: Wiley.
References 401
Edelman, G. M., and Finkel, L. (1984). In G. M. Edelman, zation of the somatosensory cortex in adult humans after
W. M. Cowan, and W. Gall, eds., Dynamic Aspects of Neo- nervous system injury. Neuroreport 5: 2593–2597.
cortical Function. 653–695. New York: Wiley.
Elbert, T., Pantev, C., Wienbruch, C., Hoke, M., Rock-
Edelman, S. (1995). Representation of similarity in three- stroh, B., and Taub, E. (1995). Increased use of the left
dimensional object discrimination. Neural Comput. 7: 408– hand in string players associated with increased cortical
423. representation of the fingers. Science 270: 305–307.
Edelman, S. (1998a). Representation is representation of Elbert, T., Sterr, A., Flor, H., Rockstroh, B., Knecht, S.,
similarity. Behav. Brain Sci. 21: 449–498. Pantev, C., Wienbruch, C., and Taub, E. (1997). Input-
increase and input-decrease types of cortical reorganization
Edelman, S. (1998b). Spanning the face space. J. Biol. Syst.
after upper extremity amputation. Exp. Brain Res. 117:
6: 265–280.
161–164.
Edelman, S. (1999). Representation and Recognition in Vision.
Elbert, T., Sterr, A., Rockstroh, B., Pantev, C., Müller,
Cambridge, MA: MIT Press.
M. M., and Taub, E. (forthcoming). Expansion of the
Edelman, S., and Bültho¤, H. (1992). Orientation depen- tonotopic area in auditory cortex of the blind.
dence in the recognition of familiar and novel views of 3D
Ellis, A. W., Young, A. W., and Flude, B. M. (1990).
objects. Vision Res. 32: 2385–2400.
Repetition priming and face recognition: Priming occurs
Edelman, S., and Duvdevani-Bar, S. (1997). A model of within the system that responds to the identity of a face.
visual recognition and categorization. Philos. Trans. R. Soc. Q. J. Exp. Psychol. 42A: 495–512.
Lond. B Biol. Sci. 352(1358): 1191–1202.
Ellis, A. W., Young, A. W., Flude, B. M., and Hay, D. C.
Edelman, S., and Intrator, N. (1997). Learning as extrac- (1987). Repetition priming of face recognition. Q. J. Exp.
tion of low-dimensional representations. In D. Medin, R. Psychol. 39A: 193–210.
Goldstone, and P. Schyns, eds., Mechanisms of Perceptual
Ellis, H. (1965). The Transfer of Learning. New York:
Learning, 353–380. San Diego: Academic Press.
Macmillan.
Edelman, S., and Intrator, N. (2000). (Coarse Coding
Ellis, H. D., Shepherd, J. W., and Davies, G. M. (1979).
of Shape Fragments) þ (Retinotopy) A Representation of
Identification of familiar and unfamiliar faces from internal
Structure. Spat. Vis. 13: 255–264.
and external features: Some implications for theories of
Edelman, S., and Weinshall, D. (1991). A self-organising face recognition. Perception 8: 431–439.
multiple-view representation of 3D objects. Biol. Cybern.
Ellison, A., and Walsh, V. (1998). Perceptual learning
64: 209–219.
in visual search: Some evidence of specificities. Vision Res.
Elberger, A. J., Smith III, E. L., and White, J. M. (1983). 38: 333–345.
Spatial dissociation of visual inputs alters the origin of the
Epstein, W. (1967). Perceptual learning resulting from ex-
corpus callosum. Neurosci. Lett. 35: 19–24.
posure to a stimulus-invariant. Am. J. Psychol. 80: 205–212.
Elbert, T. (1998). Neuromagnetism. In W. Andrä and H.
Epstein, W. (1975). Recalibration by pairing: a process of
Novak eds., Magnetism in Medicine, 190–262. New York:
perceptual learning. Perception 4: 59–72.
Wiley.
Epstein, W., Hughes, B., Schneider, S. L., and Bach-y-
Elbert, T., Candia, V., Altenmüller, E., Rau, H., Sterr, A.,
Rita, P. (1989). Perceptual learning of spatiotemporal
Rockstroh, B., Pantev, C., and Taub, E. (1998). Altera-
events: evidence from an unfamiliar modality. J. Exp. Psy-
tion of digital representations in somatosensory cortex in
chol. Hum. Percept. Perform. 15: 28–44.
focal hand dystonia. Neuroreport 9: 3571–3575.
Ergenzinger, E. R., Glasier, M. M., Hahm, J. O., and Pons,
Elbert, T., Flor, H., Birbaumer, N., Knecht, S., Hampson,
T. P. (1998). Cortically induced thalamic plasticity in the
S., Larbig, W., and Taub, E. (1994). Extensive reorgani-
primate somatosensory system. Nat. Neurosci. 1: 226–229.
402 References
Erickson, R. P. (1974). Parallel population coding in fea- Eysel, U. T., Eyding, D., and Schweigart, G. (1998). Re-
ture extraction. In F. O. Schmitt and F. G. Worden eds., petitive optical stimulation elicits fast receptive field
The Neurosciences: Third Study Program, 155–169. Cam- changes in mature visual cortex. NeuroReport 9: 949–954.
bridge, MA: MIT Press.
Eysel, U. T., Gonzalez-Aguilar, F., and Mayer, U. (1980).
Espinet, A., Almaraz, J., and Torres, P. M. (1999). Per- A functional sign of reorganization in the visual system of
ceptual learning by preschool children using stimuli with adult cats: Lateral geniculate neurons with displaced re-
varying proportions of common elements. Percept. Mot. ceptive fields after lesions of the nasal retina. Brain Res.
Skills 89: 935–942. 181: 285–300.
Esteky, H., and Tanaka, K. (1998). E¤ects of changes Eysel, U. T., Gonzalez-Aguilar, F., and Mayer, U. (1981).
in aspect ratio of stimulus shape on responses of cells in Time-dependent decrease of the extent of visual deaf-
the monkey inferotemporal cortex. Soc. Neurosci. Abstr. 24: ferentation in the lateral geniculate nucleus of adult cats
899. with small retinal lesions. Exp. Brain Res. 41: 256–263.
Estes, W. K. (1957). Of models and men. Am. Psychol. 12: Eysel, U. T., Kretschmann, U., and Schmidt-Kastner, R.
609–617. (1993). Changes of neuronal activity and immunohis-
tochemical reactions associated with photochemically in-
Etco¤, N. L., and Magee, J. J. (1992). Categorical percep-
duced thrombosis in cat visual cortex. Soc. Neurosci. Abstr.
tion of facial expressions. Cognition 44: 227–240.
19: 1668.
Eulitz, C., Eulitz, H., and Elbert, T. (1997). Di¤erential
Eysel, U. T., Muche, T., and Wörgötter, F. (1988). Lateral
outcomes from magneto- and electroencephalography for
interactions at direction selective striate neurones in the cat
the analysis of human cognition. Neurosci. Lett. 227(3):
demonstrated by local cortical inactivation. J. Physiol. 399:
185–188.
657–675.
Eurich, C. W., Dinse, H. R., Dicke, U., Godde, B., and
Eysel, U. T., and Schmidt-Kastner, R. (1991). Neuronal
Schwegler, H. (1997). Coarse coding accounts for im-
dysfunction at the border of focal lesions in cat visual cor-
provement of spatial discrimination after plastic reorgani-
tex. Neurosci. Lett. 131: 45–48.
zation in rats and humans. In Artificial Neural Networks:
Proceedings of ICANN’97, ed. W. Gerstner, A. Germond, Eysel, U. T., and Schweigart, G. (1999). Reorganization
M. Hasler, and J. D. Nicaud, 55–60. New York: Springer. of receptive fields at the border of chronic visual cortical
lesions. Cereb. Cortex 9: 101–109.
Eurich, C. W., and Schwegler, H. (1997). Coarse coding:
Calculation of the resolution achieved by a population of Eysel, U. T., Schweigart, G., Mittmann, T., Eyding, D.,
large receptive field neurons. Biol. Cybern. 76: 357–363. Qu, Y., Vandesande, F., Orban, G. A., and Arckens, L.
(1999). Reorganization in the visual cortex after retinal
Eysel, U. T. (1982). Functional reconnections without
and cortical damage. Restor. Neurol. Neurosci. 15: 153–164.
new axonal growth in a partially denervated visual relay
nucleus. Nature 299: 442–444. Eysel, U. T., Wörgötter, F., and Pape, H.-C. (1987). Local
cortical lesions abolish lateral inhibition at direction selec-
Eysel, U. T. (1997). Perilesional cortical dysfunction and
tive cells in cat visual cortex. Exp. Brain Res. 68: 606–612.
reorganization. In H. J. Freund, B. A. Sabel, and H. O.
Witte, eds., Brain Plasticity: Advances in Neurology, vol. 73, Faggin, B. M., Nguyen, K. T., and Nicolelis, M. A. (1997).
pp. 195–206. Philadelphia: Lippincott-Raven. Immediate and simultaneous sensory reorganization at cor-
tical and subcortical levels of the somatosensory system.
Eysel, U. T., Crook, J. M., and Machemer, H. F. (1990).
Proc. Natl. Acad. Sci. U S A 94: 9428–9433.
GABA-induced remote inactivation reveals cross-orienta-
tion inhibition in the cat striate cortex. Exp. Brain Res. 80: Fahle, M. (1991). A new elementary feature of vision.
626–630. Invest. Ophthalmol. Vis. Sci. 32: 2151–2155.
References 403
Fahle, M. (1994). Human pattern recognition: Parallel Ferster, D., Chung, S., and Wheat, H. (1996). Orientation
processing and perceptual learning. Perception 23: 411–427. selectivity of thalamic input to simple cells of cat visual
cortex. Nature 380: 249–252.
Fahle, M. (1997). Specificity of learning curvature, orien-
tation, and vernier discriminations. Vision Res. 37: 1885– Field, D. J. (1987). Relations between the statistics of nat-
1895. ural images and the response properties of cortical cells. J.
Opt. Soc. Am. A 4(12): 2370–2393.
Fahle, M. (1998). Orientation specificity of perceptual
learning. Invest. Ophthalmol. Vis. Sci. 39: S912. Field, D. J. (1994). What is the goal of sensory coding?
Neural Comput. 6: 559–601.
Fahle, M., and Daum, I. (1997). Visual learning and
memory as functions of age. Neuropsychologia 35: 1583– Field, D. J., and Nachmias, J. (1984). Phase reversal dis-
1589. crimination. Vision Res. 24: 333–340.
Fahle, M., and Daum, I. (Forthcoming). Perceptual learn- Field, D. J., Hayes, A., and Hess, R. F. (1993). Contour
ing in amnesic patients. Neuropsychologia (in press). integration by the human visual system: evidence for a local
‘‘association field.’’ Vision Res. 33: 173–193.
Fahle, M., and Edelman, S. (1993). Long-term learning in
vernier acuity: E¤ects of stimulus orientation, range, and Fine, I., and Jacobs, R. A. (2000). Perceptual learning
feedback. Vision Res. 33: 397–412. for a pattern discrimination task. Vision Res. 40: 3209–
3230.
Fahle, M., Edelman, S., and Poggio, T. (1995). Fast per-
ceptual learning in hyperacuity. Vision Res. 35: 3003–3013. Finnerty, G. T., Roberts, L. S., and Connors, B. W.
(1999). Sensory experience modifies the short-term
Fahle, M., and Henke-Fahle, S. (1996). Interobserver
dynamics of neocortical synapses. Nature 400: 367–371.
variance in perceptual performance and learning. Invest.
Ophthalmol. Vis. Sci. 37: 869–877. Fiorentini, A., and Berardi, N. (1980). Perceptual learning
specific for orientation and spatial frequency. Nature 287:
Fahle, M., and Morgan, M. (1996). No transfer of per-
43–44.
ceptual learning between similar stimuli in the same retinal
position. Curr. Biol. 6: 292–297. Fiorentini, A., and Berardi, N. (1981). Learning in grating
waveform discrimination: Specificity for orientation and
Fahle, M., and Skrandies, W. (1994). An electrophysio-
spatial frequency. Vision Res. 21: 1149–1158.
logical correlate of learning in motion perception. Ger. J.
Ophthalmol. 3: 427–432. Fiorentini, A., and Berardi, N. (1997). Visual perceptual
learning: A sign of neural plasticity at early stages of visual
Fantz, R. L. (1964). Visual experience in infants: Decreased
processing. Arch. Ital. Biol. 135: 157–167.
attention to familiar patterns relative to novel ones. Science
146: 668–670. Fiorentini, A., Berardi, N., Falsini, B., and Porciatti, V.
(1992). Interhemispheric transfer of visual perceptual
Farah, M. (1990). Visual Agnosia: Disorders of Object Re-
learning in callosal agenesis. Clin. Vis. Sci. 7: 133–141.
cognition and What They Can Tell Us about Normal Vision.
Cambridge, MA: MIT Press. Fischler, M. A., and Leclerc, Y. G. (1992). Recovering
3-D wire frames from line drawings. Proceedings of the
Farah, M. J., O’Reilly, R. C., and Vecera, S. P. (1993).
Image Understanding Workshop. San Francisco: Kantman.
Dissociated overt and covert recognition as an emergent
property of a lesioned neural network. Psychol. Rev. 100: Fiser, J., Biederman, I., and Cooper, E. (1996). To what
571–588 extent can matching algorithms based on direct outputs of
spatial filters account for human object recognition? Spat.
Fendick, M., and Westheimer, G. (1983). E¤ects of prac-
Vis. 10(3): 237–271.
tice and the separation of test targets on foveal and pe-
ripheral stereoacuity. Vision Res. 23: 145–150. Fisken, R. A., Garey, L. J., and Powell, T. P. S. (1975).
The intrinsic, association and commissural connections of
404 References
area 17 of the visual cortex. Philos. Trans. R. Soc. London B Foley, J. M. (1994). Human luminance pattern-vision
Biol. Sci. 272: 487–536. mechanisms: Masking experiments require a new model. J.
Opt. Soc. Am. A 11: 1710–1719.
Fitzpatrick, D. (1996). The functional organization of local
circuits in visual cortex: insights from the study of tree Foulke, E. (1991). Braille. In M. A. Heller and W. Schi¤
shrew striate cortex. Cereb. Cortex 6: 329–341. eds., The Psychology of Touch, 219–233. Hillsdale, NJ:
Erlbaum.
Flannagan, M. J., Fried, L. S., and Holyoak, K. J. (1986).
Distributional expectations and the induction of category Fox, K. (1994). The cortical component of experience-
structure. J. Exp. Psychol. Learn. Mem. Cognit. 12: 241–256. dependent synaptic plasticity in the rat barrel cortex. J.
Neurosci. 14: 7665–7679.
Flor, H., Braun, C., Elbert, T., and Birbaumer, N. (1997).
Extensive reorganization of primary somatosensory cortex Fox, K., and Daw, N. W. (1993). Do NMDA receptors
in chronic back pain patients. Neurosci. Lett. 224: 5–8. have a critical function in visual cortical plasticity? Trends
Neurosci. 16: 116–122.
Flor, H., and Elbert, T. (1998). Maladaptive consequences
of cortical reorganization in humans. Neurosci. News 1: 4– Fox, R. A., Corcoran, M., Daunton, N. G., and Morey-
11. Holton, E. (1994). E¤ects of spaceflight and hindlimb sus-
pension on the posture and gait of rats. In Vestibular and
Flor, H., Elbert, T., Knecht, S., Wienbruch, C., Pantev, C.,
Neural Front, ed. K. Taguchi, M. Igarashi, and S. Mori,
Birbaumer, N., Larbig, W., and Taub, E. (1995). Phantom-
603–606. Amsterdam: Elsevier.
limb pain as a perceptual correlate of cortical reorganization
following arm amputation. Nature 375: 482–484. Franzen, U., Lindinger, G., Lang, W., and Deecke, L.
(1991). On the functionality of the visually deprived occi-
Flor, H., Elbert, T., Muhlnickel, W., Pantev, C., Wien-
pital cortex in early blind persons. Neurosci. Lett. 124: 256–
bruch, C., and Taub, E. (1998). Cortical reorganization
259.
and phantom phenomena in congenital and traumatic
upper-extremity amputees. Exp. Brain Res. 119: 205–212. Freeman, R. B. Jr. (1966). Function of cues in the per-
ceptual learning of visual slant: An experimental and the-
Flor, H., Mühlnickel, W., Karl, A., Denke, C., Grüsser,
oretical analysis. Psychol. Monogr. 80: 1–29.
S., and Taub, E. (2000). A neural substrate for non-painful
phantom limb phenomena. NeuroReport 11: 1407–1411. Freeman, R. D., and Pettigrew, J. D. (1973). Alteration of
visual cortex from environmental asymmetries. Nature 246:
Florence, S. L., and Kaas, J. H. (1995). Large-scale reor-
359–360.
ganization at multiple levels of the somatosensory pathway
follows therapeutic amputation of the hand in monkeys. J. Frégnac, Y. (1998). Homeostasis or synaptic plasticity.
Neurosci. 15: 8083–8095. Nature 391: 845–855.
Florence, S. L., Taub, H. B., and Kaas, J. H. (1998). Large- Frégnac, Y., Bringuier, V., Chavane, F., Glaeser, L., and
scale sprouting of cortical connections after peripheral in- Lorenceau, J. (1996). An intracellular study of space and
jury in adult macaque monkeys. Science 282: 1117–1121. time representation in primary visual cortical receptive
fields. J. Physiol. Paris 90: 189–197.
Fodor, J. A., and Pylyshyn, Z. W. (1981). How direct is
visual perception? Some reflections on Gibson’s ‘‘ecological Frégnac, Y., and Imbert, M. (1978). Early development
approach.’’ Cognition 9: 139–196. of visual cortical cells in normal and dark-reared kittens:
Relationship between orientation selectivity and ocular
Földiák, P. (1990). Forming sparse representations by local
dominance. J. Physiol. 278: 27–44.
anti-Hebbian learning. Biol. Cybern. 64: 165–170.
Frégnac, Y., Shulz, D., Thorpe, S., and Bienenstock, E.
Földiák, P. (1991). Learning invariance from transformation
(1988). A cellular analogue of visual cortical plasticity.
sequences. Neural Comput. 3: 194–200.
Nature 333: 367–370.
References 405
Frégnac, Y., and Shulz, D. E. (1999). Activity-dependent Ga¤an, D. (1996). Associative and perceptual learning and
regulation of receptive field properties of cat area 17 by the concept of memory systems. Brain Res. Cogn. Brain Res.
supervised Hebbian learning. J. Neurobiol. 41: 69–82. 5: 69–80.
Frégnac, Y., Shulz, D., Thorpe, S., and Bienenstock, E. Gallant, J. L., Braun, J., and Van Essen, D. C. (1993).
(1992). Cellular analogs of visual cortical epigenesis: 1. Selectivity for polar, hyperbolic, and cartesian gratings in
Plasticity of orientation selectivity. J. Neurosci. 12: 1280– macaque visual cortex. Science 259: 100–103.
1300.
Gallant, J. L., Connor, C. E., Rakshit, S., Lewis, J. W., and
Freund, T. F., Martin, K. A. C., and Whitteridge, D. (1985). Van Essen, D. C. (1996). Neural responses to polar, hy-
Innervation of cat visual areas 17 and 18 by physiologically perbolic, and Cartesian gratings in area V4 of the macaque
identified X- and Y-type thalamic a¤erents: 1. Arborization monkey. J. Neurophysiol. 76: 2718–2739.
patterns and quantitative distribution of postsynaptic ele-
Gallistel, C. R. (1990). The Organization of Learning. Cam-
ments. J. Comp. Neurol. 242: 263–274.
bridge, MA: MIT Press.
Fried, I., MacDonald, K. A., and Wilson, C. L. (1997).
Galper, R. E. (1970). Recognition of faces in photo-
Single neuron activity in human hippocampus and amyg-
graphic negative. Psychonom. Sci. 19: 207–208.
dala during recognition of faces and objects. Neuron 18:
753–765. Galton, F. (1883). Inquiries into Human Faculty and Its De-
velopment. London: Macmillan.
Friedman, J. (1994). Flexible Metric Nearest Neighbor Classi-
fication. Technical report. Stanford University. Galuske, R. A. W., and Singer, W. (1996). The origin and
topography of long-range intrinsic projections in cat visual
Fuhr, P., Cohen, L. G., Dang, N., Findley, T. W.,
cortex: A developmental study. Cereb. Cortex 6: 417–
Haghighi, S., Oro, J., and Hallett, M. (1992). Physiologi-
430.
cal analysis of motor reorganization following lower
limb amputation. Electroenc. Clin. Neurophysiol. 85: 53– Gardner, D. (1993). Static determinants of synaptic
60. strength. In Gardner, ed., The Neurobiology of Neural Net-
works. Cambridge, MA: MIT Press.
Fujita, I., Tanaka, K., Ito, M., and Cheng, K. (1992).
Columns for visual features of objects in monkey infero- Gardner, E. P., and Costanzo, R. M. (1980). Temporal
temporal cortex. Nature 360: 343–346. integration of multiple-point stimuli in primary somato-
sensory cortical receptive fields of alert monkeys. J. Neu-
Furmanski, C. S., and Engel, S. A. (2000). Perceptual
rophysiol. 43: 444–468.
learning in object recognition: Object specificity and size
invariance. Vision Res. 40: 473–484. Garraghty, P. E., and Kaas, J. H. (1991). Functional reor-
ganization in adult monkey thalamus after peripheral nerve
Fuster, J. M., and Jervey, J. P. (1981). Inferotemporal
injury. NeuroReport 2: 747–750.
neurons distinguish and retain behaviorally relevant fea-
tures of visual stimuli. Science 212: 952–955. Garraghty, P. E., and Kaas, J. H. (1992). Dynamic features
of sensory and motor maps. Curr. Opin. Neurobiol. 2: 522–
Fyfe, C., and Baddeley, R. (1995). Finding compact and
527.
sparse-distributed representations of visual images. Network
6: 333–344. Gauthier, I., and Tarr, M. J. (1997). Becoming a ‘‘greeble’’
expert: Exploring mechanisms for face recognition. Vision
Gabor, D. (1946). Theory of communication. J. Inst. Elect.
Res. 37: 1673–1682.
Eng. Lond. 93: 429–457.
Gauthier, I., Tarr, M. J., Anderson, A. W., Skudlarski, P.,
Gabrieli, J. D., Milberg, W., Keane, M. M., and Corkin,
and Gore, J. C. (1999). Activation of the middle fusiform
S. (1990). Intact priming of patterns despite impaired
‘‘face area’’ increases with expertise in recognizing novel
memory. Neuropsychologia 28: 417–427.
objects. Nat. Neurosci. 2: 568–573.
406 References
Geisler, W. S., and Albrecht, D. G. (1992). Cortical neu- Gibson, E., Owsley, C., and Johnston, J. (1978). Percep-
rons: Isolation of contrast gain control. Vision Res. 32: tion of invariants by five-month-old infants. Dev. Psychol.
1409–1410. 14: 407–415.
Geisler, W. S., and Albrecht, D. G. (1997). Visual cortex Gibson, J. J., and Gibson, E. J. (1955). Perceptual learning:
neurons in monkeys and cats: Detection, discrimination Di¤erentiation or enrichment? Psychol. Rev. 62: 32–41.
and identification. Vis. Neurosci. 14: 897–919.
Gick, M., and Holyoak, K. (1980). Analogical problem
Geldard, F. A., and Sherrick, C. E. (1972). The cutaneous solving. Cogn. Psychol. 12: 306–355
‘‘rabbit’’: A perceptual illusion. Science 178: 178–179.
Gilbert, C. D. (1983). Microcircuitry of the visual cortex.
Gellatly, A. R. (1982). Perceptual learning of illusory Annu. Rev. Neurosci. 6: 217–247.
contours and colour. Perception 11: 655–661.
Gilbert, C. D. (1992). Horizontal integration and cortical
Georgopoulos, A. P., Schwartz, A. B., and Kettner, R. E. dynamics. Neuron 9: 1–13.
(1986). Neural population coding of movement direction.
Gilbert, C. D. (1993). Circuitry, architecture and functional
Science 233: 1416–1419.
dynamics of visual cortex. Cerebr. Cortex 3: 373–386.
Gerrits, H. J. M., and Timmermann, G. J. M. E. (1969).
Gilbert, C. D. (1994). Early perceptual learning. Proc. Natl.
The filling-in process in patients with retinal scotomata.
Acad. Sci. U S A 91: 1195–1197.
Vision Res. 9: 439–442.
Gilbert, C. D. (1994). Neuronal dynamics and perceptual
Gerrits, H. J. M., de Haan, B., Vendrick, A. J. H. (1966).
learning. Curr. Biol. 4: 627–629.
Experiments with retinal stabilized images: Relations be-
tween the observations and neural data. Vision Res. 6: 427– Gilbert, C. D. (1998). Adult cortical dynamics. Physiol.
440. Rev. 78: 467–485.
Gerrits, H. J. M., and Vendrik, A. J. H. (1970). Simulta- Gilbert, C. D., and Wiesel, T. N. (1979). Morphology and
neous contrast, filling-in process and information process- intracortical projections of functionally characterised neu-
ing in man’s visual system. Exp. Brain Res. 11: 411–430. rones in the cat visual cortex. Nature 280: 120–125.
Ghazanfar, A. A., Stambaugh, C. R., and Nicolelis, M. A. Gilbert, C. D., and Wiesel, T. N. (1983). Clustered intrinsic
(2000). Encoding of tactile stimulus location by somato- connections in cat visual cortex. J. Neurosci. 3: 1116–1133.
sensory thalamocortical ensembles. J. Neurosci. 20: 3761– Gilbert, C. D., and Wiesel, T. N. (1985). Intrinsic con-
3775. nectivity and receptive field properties in visual cortex.
Gibson, E. J. (1969). Principles of Perceptual Learning and Vision Res. 25: 365–374.
Development. New York: Appleton-Century-Crofts. Gilbert, C. D., and Wiesel, T. N. (1989). Columnar spe-
Gibson, E. J. (1941). Retroactive inhibition as a function cificity of intrinsic horizontal and corticocortical con-
of the degree of generalization between tasks. J. Exp. Psy- nections in cat visual cortex. J. Neurosci. 9: 2432–2442.
chol. 28: 93–115. Gilbert, C. D., and Wiesel, T. N. (1992). Receptive field
Gibson, E. J. (1953). Improvement in perceptual judge- dynamics in adult primary visual cortex. Nature 356: 150–
ments as a function of controlled practice or training. Psy- 152.
chol. B 50: 401–431. Gilbert, C. D., Das, A., Ito, M., Kapadia, M., and West-
Gibson, E. J. (1963). Perceptual learning. Annu. Rev. Psy- heimer, G. (1996). Spatial integration and cortical dy-
chol. 14: 29–56. namics. Proc. Natl. Acad. Sci. U S A 93: 615–622.
Gibson, E. J., and Walk, R. D. (1956). The e¤ect of pro- Gilbert, C. D., Ito, M., Kapadia, M., and Westheimer, G.
longed exposure to visually presented patterns on learning (2000). Interactions between attention, context and learn-
to discriminate them. J. Comp. Physiol. Psychol. 49: 239–242. ing in primary visual cortex. Vision Res. 40: 1217–1226.
References 407
Gilbert, C. D., Sigman, M., and Crist, R. E. (2001). The Goodale, M. A., and Humphrey, G. K. (1998). The objects
neural basis of perceptual learning. Neuron 31: 681–697. of action and perception. Cognition 67: 181–207.
Gilbert, D. K., and Rogers, W. A. (1996). Age-related dif- Goodale, M., and Milner, A. (1992). Separate visual path-
ferences in perceptual learning. Hum. Factors 38: 417–424. ways for perception and action. Trends Neurosci. 15: 20–
25.
Girosi, F., Jones, M., and Poggio, T. (1995). Regulariza-
tion theory and neural networks architectures. Neural Goodall, M. C. (1960). Performance of stochastic nets.
Comput. 7: 219–269. Nature 185: 557–558.
Gliner, J. A., Mihevic, P. M., and Horvath, S. M. (1983). Goren, C. C., Sarty, M., and Wu, R. W. K. (1975). Visual
Spectral analysis of electroencephalogram during percep- following and pattern discrimination of face-like stimuli by
tual-motor learning. Biol. Psychol. 16: 1–13. new-born infants. Pediatrics 56: 544–549.
Gluck, M. A., and Granger, R. (1993). Computational Gottlieb, G. L., Corcos, D. M., Jaric, S., and Agarwal, G.
models of the neural bases of learning and memory. Annu. C. (1988). Practice improves even the simplest movements.
Rev. Neurosci. 16: 667–706. Exp. Brain Res. 73: 436–440.
Godde, B., Spengler, F., and Dinse, H. R. (1996). Asso- Grady, C. L., and Craik, F. I. M. (2000). Changes in
ciative pairing of tactile stimulation induces somatosensory memory processing with age. Curr. Opin. Neurobiol. 10:
cortical reorganization in rats and humans. NeuroReport 8: 224–231.
281–285.
Grady, C. L., Horwitz, B., Pietrini, P., Mentis, M. J.,
Godde, B., Stau¤enberg, B., Spengler, F., and Dinse, H. Ungerleider, L. G., Rapoport, S. I., and Haxby, J. V.
R. (2000). Tactile coactivation induced changes in (1996). E¤ect of task di‰culty on cerebral blood flow
spatial discrimination performance. J. Neurosci. 20: 1597– during perceptual matching of faces. Hum. Brain Mapp.
1604. 4074: 227–239.
Godecke, I., and Bonhoe¤er, T. (1996). Development of Graham, D. B., and Allison, N. M. (1998). Characterising
identical orientation maps for two eyes without common virtual eigensignatures for general purpose face recognition.
visual experience. Nature 379: 251–254. In H. Wechsler et al., eds., Face Recognition: From Theory to
Applications, pp. 446–456. Berlin: Springer.
Gold, J., Bennett, P. J., and Sekuler, A. B. (1999). Signal
but not noise changes with perceptual learning. Nature Graham, S. (1999). Retrospective revaluation and inhibi-
402: 176–178. tory associations: does perceptual learning modulate our
perception of the contingencies between events? Q. J.
Goldstein, B. E. (1996). Sensation and Perception. Pacific
Exp. Psychol. B. 52: 159–185.
Grove, CA: Brooks/Cole.
Granger, R., and Lynch, G. (1991). Higher olfactory pro-
Goldstone, R. L. (1998). Perceptual learning. Annu. Rev.
cesses: perceptual learning and memory. Curr. Opin. Neu-
Psychol. 49: 585–612.
robiol. 1: 209–214.
Goldstone, R. L., and Barsalou, L. W. (1998). Reuniting
Granger, R., Whitson, J., Larson, J., and Lynch, G. (1994).
perception and cognition: The perceptual bases of similarity
Non-Hebbian properties of long-term potentiation enable
and rules. Cognition 65: 231–262.
high-capacity encoding of temporal sequences. Proc. Natl.
Gollin, E. S. (1960). Developmental studies of visual rec- Acad. Sci. U S A 91: 10104–10108.
ognition of incomplete objects. Percept. Mot. Skills 11: 289–
Gray, C. M., König, P., Engel, A. K., and Singer, W.
298.
(1989). Oscillatory responses in cat visual cortex exhibit
Gollin, E. S. (1965). Perceptual learning of incomplete inter-columnar synchronization which reflects global stim-
pictures. Percept. Mot. Skills 21: 439–445. ulus parameters. Nature 338: 334–337.
408 References
Green, J. B., Sora, E., Bialy, Y., Ricamato, A., and Grunke, M. E., and Pisoni, D. B. (1982). Some experiments
Thatcher, R. W. (1998). Cortical sensorimotor reorgani- on perceptual learning of mirror-image acoustic patterns.
zation after spinal cord injury. Neurology 50: 1115–1121. Percept. Psychophys. 31: 210–218.
Greenspan, S. L., Nusbaum, H. C., and Pisoni, D. B. Gu, X., and Fortier, P. A. (1996). Early enhancement but
(1988). Perceptual learning of synthetic speech produced no late changes of motor responses induced by intracortical
by rule. J. Exp. Psychol. Learn. Mem. Cogn. 14: 421–433. microstimulation in the ketamine-anesthetized rat. Exp.
Brain. Res. 108: 119–128.
Gregory, R. (1972). Eye and Brain: The Psychology of Seeing.
2d ed. London: Weidenfeld and Nicolson. Gunderson, V. M., and Sackett, G. P. (1984). Develop-
ment of pattern recognition in infant pigtailed macaques
Gregory, R., and Wallace, J. (1963). Recovery from Early
(Macaca nemestrina). Dev. Psychol. 22: 477–480.
Blindness: A Case Study. Experimental Psychology Society
Monograph 2. Cambridge: W. He¤er. Guttman, N. (1963). Laws of behavior and facts of per-
ception. In S. Koch, ed., Psychology: A study of a Science,
Greuel, J. M., Luhmann, H. J., and Singer, W. (1988).
vol. 5, pp. 114–178. New York: McGraw-Hill.
Pharmacological induction of use-dependent receptive
field modifications in the visual cortex. Science 242: 74–77. Guzman, A. (1971). Analysis of curved line drawings using
context and global information. Mach. Intell. 6: 325–375.
Gross, C. G. (1973). Visual functions of inferotemporal
cortex. In Handbook of Sensory Physiology, vol. 7, part 3B, Hagemann, G., Redecker, C., Neumann-Haefelin, T.,
ed. R. Jung, 451–482. Berlin: Springer. Freund, H.-J., and Witte, O. W. (1998). Increased long-
term potentiation in the surround of experimentally in-
Gross, C. G. (1992). Representation of visual stimuli in in-
duced focal cortical infarction. Ann. Neurol. 44: 255–258.
ferior temporal cortex. Philos. Trans. R. Soc. Lond. B Biol.
Sci. 335: 3–10. Hall, E. J., Flament, D., Fraser, C., and Lemon, R. N.
(1990). Non-invasive brain stimulation reveals reorganized
Gross, C. G. (1994). How inferior temporal cortex became
cortical outputs in amputees. Neurosci. Lett. 116: 379–386.
a visual area. Cereb. Cortex 4: 455–469.
Hall, G. (1991). Perceptual and Associative Learning. Oxford
Gross, C. G., Bender, D. B., and Rocha-Miranda, C. E.
Psychology Series No.18. Oxford, UK: Clarendon Press.
(1969). Visual receptive fields of neurons in inferotemporal
cortex of the monkey. Science 166: 1303–1306. Halligan, P. W., Marshall, J. C., Wade, D. T., Davey, J.,
and Morrison, D. (1993). Thumb in cheek? Sensory reor-
Gross, C. G., Rocha-Miranda, C. E., and Bender, D. B.
ganization and perceptual plasticity after limb amputation.
(1972). Visual properties of neurons in inferotemporal
NeuroReport 4: 233–236.
cortex of the macaque. J. Neurophysiol. 35: 96–111.
Hancock, P. J. B., Baddeley, R. J., and Smith, L. S. (1992).
Grossberg, S. (1987). Competitive learning: From interac-
The principal components of natural images. Network 3:
tive activation to adaptive resonance. Cogn. Sci. 11: 23–63.
61–72.
Grossberg, S., and Mingolla, E. (1985). Neural dynamics
Hancock, P. J. B., Bruce, V., and Burton, A. M. (1998). A
of perceptual grouping: Textures, boundaries, and emer-
comparison of two computer-based face identification
gent segmentations. Percep. Psychophys. 38: 141–171.
systems with human perceptions of faces. Vision Res. 38:
Grossberg, S., and Williamson, J. R. (2001). A neural 2277–2288.
model of how horizontal and interlaminar connections of
Hancock, P. J. B., Burton, A. M., and Bruce, V. (1996).
visual cortex develop into adult circuits that carry out
Face processing: Human perception and principal compo-
perceptual grouping and learning. Cereb. Cortex 11: 37–58.
nents analysis. Mem. Cognit. 24: 26–40.
Gruber, H. E. (1995). Insight and a¤ect in the history of
Haralick, R. M. (1980). Edge and region analysis for digi-
science. In The Nature of Insight, ed. R. J. Sternberg and
tal image data. Comput. Graph. Image Proc. 12: 60–73.
J. E. Davidson. Cambridge, MA: MIT Press.
References 409
Harnad, S., ed. (1987). Categorical Perception: The Ground- Helmholtz, H. von (1866/1911). Helmholtz’s Physiological
work of Cognition. New York: Cambridge University Press. Optics. 3d ed., trans. and ed. J. P. Southwell. Rochester,
NY: Optical Society of America.
Harpur, G. F., and Prager, R. W. (1996). Development of
low entropy coding in a recurrent network. Network 7: Hering, E. (1861). Beiträge zur Physiologie: Zur Lehre vom
277–284. Ortsinne der Netzhaut. Leipzig: Engelmann.
Harpur, J. G., Estabrooks, K. A., Allen, N. J., and Asaph, Hershberger, W. (1970). Attached-shadow orientation
C. A. (1978). Perceptual versus mediational learning in a perceived as depth by chickens reared in an environment
total change concept-shift paradigm. Percept. Mot. Skills 46: illuminated from below. J. Comp. Physiol. Psychol. 73:
563–569. 407–411.
Harris, C. (1963). Adaptation to displaced vision: Visual, Hertz, J., Krogh, A., and Palmer, R. G. (1991). Introduction
motor, or proprioceptive change? Science 140: 812–813. to the Theory of Neural Computation. Redwood City, CA:
Addison-Wesley and Santa Fe Institute.
Harris, J. A., and Diamond, M. E. (2000). Ipsilateral and
contralateral transfer of tactile learning. NeuroReport 11: Herzog, M. H., and Fahle, M. (1994). Learning without
263–266. attention? In N. Elsner and H. Breer, eds., Proceedings of the
Twenty-second Göttingen Neurobiology Conference, 1994, vol.
Hartline, H. K. (1940). The receptive fields of optic nerve
II, no. 817. Stuttgart: Thieme.
fibers. Am. J. Physiol. 130: 690–699.
Herzog, M. H., and Fahle, M. (1997). The role of feed-
Hasselmo, M. E., Rolls, E. T., and Baylis, G. C. (1989).
back in learning a vernier discrimination task. Vision Res.
The role of expression and identity in the face-selective
37: 2133–2141.
responses of neurons in the temporal visual cortex of the
monkey. Behav. Brain Res. 32: 203–218. Herzog, M. H., and Fahle, M. (1998). Modeling percep-
tual learning: Di‰culties and how they can be overcome.
Hata, Y., Tsumoto, T., Sato, H., and Tamura, H. (1991).
Biol. Cybern. 78: 107–117.
Horizontal interactions between visual cortical neurones
studied by cross-correlation analysis in the cat. J. Physiol. Herzog, M. H., and Fahle, M. (1999). E¤ects of biased
(Lond.) 441: 593–614. feedback on learning and deciding in a vernier discrimi-
nation task. Vision Res. 39: 4232–4243.
Hatfield, G., and Epstein, W. (1985). The status of the
minimum principle in the theoretical analysis of vision. Herzog, M. H., Broos, A. H., and Fahle, M. (1999). Prac-
Psychol. Bull. 97: 155–186. ticing a vernier discrimination task with non-uniformly
distributed stimuli influences decision but not learning
Haykin, S., ed. (1994). Blind Deconvolution. Englewood
processes. Invest. Ophthalmol. Vis. Sci. 40: 3077.
Cli¤s, NJ: Prentice-Hall.
Hietanen, J. K., Perrett, D. I., Oram, M. W., Benson,
Hebb, O. D. (1949). The Organization of Behavior. A Neu-
P. J., and Dittrich, W. H. (1992). The e¤ects of lighting
ropsychological Theory. New York: Wiley.
conditions on responses of cells selective for face views
Heeger, D. J. (1992). Normalization of cell responses in cat in the macaque temporal cortex. Exp. Brain Res. 89: 157–
striate cortex. Vis. Neurosci. 9: 181–197. 171.
Heinen, S. J., and Skavenski, A. A. (1991). Recovery of Hikosaka, O., Miyauchi, S., and Shimojo, S. (1993). Visual
visual responses in foveal V1 neurons following bilateral attention revealed by an illusion of motion. Neurosci. Res.
foveal lesions in adult monkey. Exp. Brain Res. 83: 670–674. 18: 11–18.
Held, R. (1999). Visual development in infants. In G. Hikosaka, O., Miyauchi, S., and Shimojo, S. (1996). Ori-
Adelman and B. H. Smith, eds., Encyclopedia of Neuroscience, enting a spatial attention: Its reflexive, compensatory, and
2nd ed., vol. 2., pp. 2124–2127. New York: Elsevier. voluntary mechanisms. Cogn. Brain Res. 5: 1–9.
410 References
Hill, H., and Bruce, V. (1996). E¤ects of lighting on Honey, R. C., and Bateson, P. (1996). Stimulus compari-
matching facial surfaces. J. Exp. Psychol. Hum. Percept. Per- son and perceptual learning: Further evidence and evalua-
form. 22: 986–1004. tion from an imprinting procedure. Q. J. Exp. Psychol. B.
49: 259–269.
Hinton, G. E. (1989). Connectionist learning procedures.
Artif. Intell. 40: 185–234. Honey, R. C., Bateson, P., and Horn, G. (1994). The role
of stimulus comparison in perceptual learning: An investi-
Hinton, G. E., McClelland, J. L., and Rumelhart, D. E.
gation with the domestic chick. Q. J. Exp. Psychol. B. 47:
(1986). Distributed representations. In J. A. Feldman, P. J.
83–103.
Hayes, and D. E. Rumelhart, eds., Parallel Distributed Pro-
cessing: Exploration in the Microstructure of Cognition. Vol. 1, Honig, M. G., and Hume, R. I. (1989). DiI and diO:
Foundations. 77–109. Cambridge, MA: MIT Press. Versatile fluorescent dyes for neuronal labelling and path-
way tracing. Trends Neurosci. 12: 333–341.
Hinton, G., Williams, C., and Revow, M. (1992). Adap-
tive elastic models for hand-printed character recognition. Horel, J. A., Pytko-Joiner, D. E., Voytko, M. L., and
In J. Moody, S. Hanson, and R. Lippman, eds., Advances in Salsbury, K. (1987). The performance of visual tasks while
Neural Information-Processing Systems, vol. 4, pp. 512–519. segments of the inferotemporal cortex are suppressed by
San Meteo, CA: Morgan Kaufmann. cold. Behav. Brain Res. 23: 29–42.
Hintzman, D. L. (1994). Twenty-five years of learning and Horn, B. K. P. (1975). Obtaining shape from shading in-
memory: Was the cognitive revolution a mistake? In C. formation. In The Psychology of Computer Vision, ed. P. H.
Umiltà and M. Moscovitch, eds., Attention and Performance Winston. New York: McGraw-Hill.
XV, 360–391. Cambridge, MA: MIT Press.
Houzel, J.-C., Milleret, C., and Innocenti, G. (1994).
Hirsch, J. A., and Gilbert, C. D. (1993). Long-term changes Morphology of callosal axons interconnecting areas 17 and
in synaptic strength along specific intrinsic pathways in the 18 of the cat. Eur. J. Neurosci. 6: 898–917.
cat visual cortex. J. Physiol. 461: 247–262.
Howard, I. P. (1971). Perceptual learning and adaptation.
Hochberg, J. E., and Brooks, V. (1958). E¤ects of pre- Br. Med. Bull. 27: 248–252.
viously associated annoying stimuli (auditory) on visual
Hubel, D. H. (1995). Eye, Brain, and Vision. New York:
recognition thresholds. J. Exp. Psychol. 55: 490–491.
Scientific American Library.
Hochmair-Desoyer, I. J., and Burian, K. (1985). Reim-
Hubel, D. H., and Wiesel, T. N. (1959). Receptive fields
plantation of a modulated scala tympani electrode: Impact
of single neurones in the cat’s striate cortex. J. Physiol. 148:
on psychophysical and speech discrimination abilities. Ann.
574–591.
Otol. Rhinol. Laryngol. 94: 65–70.
Hubel, D. H., and Wiesel, T. N. (1962). Receptive fields,
Hochstein, S., Lobovsky, S., Laiwand, R., and Ahissar, M.
binocular interaction and functional architecture in the cat’s
(2000). Dual-task performance within and across dimen-
visual cortex. J. Physiol. 160: 106–154.
sions for spatially overlapping and non-overlapping tasks.
Invest. Ophthalmol. Vis. Sci. 41(4): 1048. Hubel, D. H., and Wiesel, T. N. (1965). Binocular inter-
action in striate cortex of kittens reared with artificial
Hock, H. S., Webb, E., and Cavedo, L. C. (1987). Per-
squint. J. Neurophysiol. 28: 1041–1059.
ceptual learning in visual category acquisition. Mem. Cognit.
15: 544–556. Hubel, D. H., and Wiesel, T. N. (1967). Cortical and cal-
losal connections concerned with the vertical meridian of
Ho¤man, W. (1966). The lie algebra of visual perception.
visual fields in the cat. J. Neurophysiol. 30: 1561–1573.
J. Math. Psychol. 3: 65–98.
Hubel, D. H., and Wiesel, T. N. (1968). Receptive fields
Hollins, M. (1989). Understanding Blindness. Hillsdale, NJ:
and functional architecture of monkey striate cortex. J.
Erlbaum.
Physiol. (Lond.) 195: 215–243.
References 411
Hubel, D. H., and Wiesel, T. N. (1970). The period of Hurlbert, A. (2000). Visual perception: Learning to see
susceptibility to the physiological e¤ects of unilateral eye through noise. Curr. Biol. 10: R231–233.
closure in kittens. J. Physiol. 206: 419–436.
Imbert, M., and Buisseret, P. (1975). Receptive field
Hubel, D. H., and Wiesel, T. N. (1974). Uniformity of characteristics and plastic properties of visual cortical cells
monkey striate cortex: A parallel relationship between field in kittens reared with or without visual experience. Exp.
size, scatter, and magnification factor. J. Comp. Neurol. 158: Brain Res. 22: 25–36.
295–306.
Ingram, D. K. (1988). Motor performance variability dur-
Hubel, D. H., and Wiesel, T. N. (1977). Functional ar- ing aging in rodents. Ann. N. Y. Acad. Sci. 515: 70–95.
chitecture of macaque monkey visual cortex. Proc. R. Soc.
Innocenti, G. M. (1986). Postnatal development of corti-
Lond. B. 198: 1–59.
cocortical connections. Ital. J. Neurol. Sci. 5: 25–28.
Hubel, D. H., and Wiesel, T. N. (1998). Early exploration
Innocenti, G. M., and Caminiti, R. (1980). Postnatal
of the visual cortex. Neuron 20: 401–412.
shaping of callosal connections from sensory areas. Exp.
Hubel, D. H., Wiesel, T. N., and LeVay, S. (1977). Plas- Brain Res. 38: 381–394.
ticity of ocular dominance columns in monkey striate cor-
Innocenti, G. M., and Frost, D. O. (1979). E¤ects of visual
tex. Philos. Trans. R. Soc. Lond. B Biol. Sci. 278: 377–409.
experience on the maturation of the e¤erent system to the
Huber, P. J. (1985). Projection pursuit (with discussion). corpus callosum. Nature 280: 231–234.
Ann. Stat. 13: 435–475.
Innocenti, G. M., Fiore, L., and Caminiti, R. (1977). Exu-
Hubscher, C. H., and Johnson, R. D. (1999). Changes in berant projection into the corpus callosum from the visual
neuronal receptive field characteristics in caudal brain stem cortex of newborn cats. Neurosci. Lett. 4: 237–242.
following chronic spinal cord injury. J. Neurotrauma 16:
Intrator, N. (1992). Feature extraction using an unsu-
533–541.
pervised neural network. Neural Comput. 4: 98–107.
Hughes, B., Epstein, W., Schneider, S., and Dudock, A.
Intrator, N. (1993). Combining exploratory projection
(1990). An asymmetry in transmodal perceptual learning.
pursuit and projection pursuit regression. Neural Comput.
Percept. Psychophys. 48: 143–150.
5: 443–455.
Hume, D. (1738/1956). A Treatise of Human Nature.
Intrator, N., and Edelman, S. (1997). Learning low-
London: Everyman’s Library.
dimensional representations of visual objects with exten-
Humphrey, N. K., and Keeble, G. R. (1976). How mon- sive use of prior knowledge. Network 8: 259–281.
keys acquire a new way of seeing. Perception 5: 51–56.
Ippolito, M. F., and Tweney, R. (1995). The inception of
Humphreys, G. W., and Riddoch, M. J. (1987). To See but insight. In The Nature of Insight, ed. R. J. Sternberg and
Not to See: A Case Study of Visual Agnosia. Hillsdale, NJ: J. E. Davidson. Cambridge, MA: MIT Press.
Erlbaum.
Irvine, D. R. F., and Rajan, R. (1996). Injury- and use-
Huntley, G. W. (1997). Correlation between patterns of related plasticity in the primary sensory cortex of adult
horizontal connectivity and the extent of short-term rep- mammals: possible relationship to perceptual learning.
resentational plasticity in rat motor cortex. Cereb. Cortex 7: Clin. Exp. Pharmacol. Physiol. 23: 939–947.
143–156.
Irvine, D. R., Martin, R. L., Klimkeit, E., and Smith, R.
Hupé, J. M., James, A. C., Payne, B. R., Lomber, S. G., (2000). Specificity of perceptual learning in a fre-
Girard, P., and Bullier, J. (1998). Cortical feedback quency discrimination task. J. Acoust. Soc. Am. 108: 2964–
improves discrimination between figure and background 2968.
by V1, V2, and V3 neurons. Nature 394: 784–787.
Ishai, A., and Sagi, D. (1995). Common mechanisms of
visual imagery and perception. Science 268: 1772–1774.
412 References
Ito, M., Fujita, I., Tamura, H., and Tanaka, K. (1994). Johnson, M. H., Dziurawiec, S., Ellis, H., and Morton, J.
Processing of contrast polarity of visual images in infero- (1991). Newborns’ preferential tracking of face-like stimuli
temporal cortex of the macaque monkey. Cereb. Cortex 5: and its subsequent decline. Cognition 40: 1–19.
499–508.
Johnston, A., Hill, H., and Carman, N. (1992). Recog-
Ito, M., Tamura, H., Fujita, I., and Tanaka, K. (1995). Size nising faces: E¤ects of lighting direction, inversion and
and position invariance of neuronal responses in monkey brightness reversal. Perception 21: 365–375.
inferotemporal cortex. J. Neurophysiol. 73: 218–226.
Johnston, R. A., and Ellis, H. D. (1995). Age e¤ects in the
Ito, M., Westheimer, G., and Gilbert, C. D. (1998). At- processing of typical and distinctive faces. Q. J. Exp. Psy-
tention and perceptual learning modulate contextual in- chol. 48: 447–465.
fluences on visual perception. Neuron 20: 1191–1197.
Johnston, R. A., Milne, A. B., Williams, C., and Hosie, J.
Jacoby, L. L., and Dallas, M. (1981). On the relationship (1997). Do distinctive faces come from outer space? An
between autobiographical memory and perceptual learn- investigation of the status of a multidimensional face-space.
ing. J. Exp. Psychol. Gen. 110: 306–340. Vis. Cognit. 4: 1–112.
Jain, N., Catania, K. C., and Kaas, J. H. (1997). Deactivi- Jolicoeur, P. (1990). Orientation congruency e¤ects on the
ation and reactivation of somatosensory cortex after dorsal identification of disoriented shapes. J. Exp. Psychol. Hum.
spinal cord injury. Nature 368: 495–498. Percept. Perform. 16: 351–364.
Jain, N., Florence, S. L., Qi, H. X., and Kaas, J. H. (2000). Jones, E. G., and Pons, T. P. (1998). Thalamic and brain-
Growth of new brainstem connections in adult monkeys stem contributions to large-scale plasticity of primate so-
with massive sensory loss. Proc. Natl. Acad. Sci. U S A 97: matosensory cortex. Science 282: 1121–1125.
5546–5550.
Jones, M. J., and Poggio, T. A. (1995). Model-based
James, W. (1890/1950). Principles of Psychology. Vol. 1. New matching by linear combinations of prototypes. In Proceed-
York: Dover. ings of the Fifth International Conference on Computer Vision,
Los Alamitos, CA: IEEE Press.
Jancke, J., Erlhagen, W., Dinse, H. R., Akhavan, A. C.,
Giese, M., Steinhage, A., and Schöner, G. (1999). Para- Jones, M. J., Sinha, P., Vetter, T., and Poggio, T. (1997).
metric population representation of retinal location: Neu- Top-down learning of low-level vision tasks. Curr. Biol. 7:
ronal interaction dynamics in cat primary visual cortex. J. 991–994.
Neurosci. 19: 9016–9028.
Joseph, J. S., Chun, M. M., and Nakayama, K. (1997).
Jenkins, W. M., and Merzenich, M. M. (1987). Reorga- Attentional requirements in a ‘preattentive’ feature search
nization of neocortical representations after brain injury: A task. Nature 387: 805–807.
neurophysiological model of the bases of recovery from
Joublin, F., Spengler, F., Wacquant, S., and Dinse, H. R.
stroke. Prog. Brain Res. 71: 249–266.
(1996). A columnar model of somatosensory reorganiza-
Jenkins, W. M., Merzenich, M. M., Ochs, M. T., Allard, tional plasticity based on Hebbian and non-Hebbian
T., and Guic-Robles, E. (1990). Functional reorganization learning rules. Biol. Cybern. 74: 275–286.
of primary somatosensory cortex in adult owl monkeys
Juettner, M., Caelli, T., and Rentschler, I. I. (1997).
after behaviorally controlled tactile stimulation. J. Neuro-
Evidence-based pattern classification: A structural approach
physiol. 63: 82–104.
to human perceptual learning and generalization. J. Math.
Jensen, A. R. (1966). Social class and perceptual learning. Psychol. 41: 244–259.
Ment. Hyg. 50: 226–239.
Julesz, B. (1971). Foundations of Cyclopean Perception. Uni-
Johansson, G. (1973). Visual perception of biological versity of Chicago Press.
motion and a model of its analysis. Percept. Psychophys. 14:
Julesz, B. (1981). Textons: The elements of texture per-
201–211.
ception and their interactions. Nature 290: 91–97.
References 413
Julesz, B. (1986). Texton gradients: The texton theory Kaas, J. H., Jain, N., and Florence, S. L. (1998). The re-
revisited. Biol. Cybern. 54: 245–251. activation of sensory cortex after deactivation by peripheral
nerve or spinal cord injury. Neurosci. News 1: 12–17.
Julesz, B. (1990). Early vision is bottom-up except for focal
attention. Cold Spring Harb. Symp. Quant. Biol. 55: 973–978. Kaas, J. H., Krubitzer, L. A., Chino, Y. M., Langston, A.
L., Polley, E. H., and Blair, N. (1990). Reorganization of
Jürgens, M., and Dinse, H. R. (1995). Spatial and temporal
retinotopic cortical maps in adult mammals after lesions of
integration properties of cortical somatosensory neurons in
the retina. Science 248: 229–231.
aged rats: Lack of age-related cortical changes in behav-
iorally unimpaired individuals of high age. Soc. Neurosci. Kalarickal, G. J., and Marshall, J. A. (1999). Models of
Abstr. 21: 197. receptive-field dynamics in visual cortex. Vis. Neurosci. 16:
1055–1081.
Jürgens, M., and Dinse, H. R. (1997a). Use-dependent
plasticity of SI cortical hindpaw neurons induced by Kanade, T. (1981). Recovery of the three-dimensional
modification of walking in adult rats: A model for age re- shape of an object from a single view. Artif. Intell. 17(1–3):
lated alterations. Soc. Neurosci. Abstr. 23: 1800. 409–460.
Jürgens, M., and Dinse, H. R. (1997b). Di¤erential e¤ects Kandel, E. R., Schwartz, J. H., and Jessel, T. M. (1992).
of the Ca 2þ -influxblocker nimodipine on receptive field Principles of Neural Science, 3rd ed. New York: Elsevier and
properties and response latencies of somatosensory cortical North-Holland.
neurons in aged rats. In Internal Report 96-10, 1–23. Insti-
Kanizsa, G. (1979). Organization in Vision. New York:
tut für Neuroinformatik, Ruhr-University, Bochum.
Praeger.
Jüttner, M., and Rentschler, I. (1996). Reduced perceptual
Kapadia, M. K., Ito, M., Gilbert, C. D., and Westheimer,
dimensionality in extrafoveal vision. Vision Res. 36: 1007–
G. (1995). Improvement in visual sensitivity by changes in
1022.
local context: Parallel studies in human observes and in V1
Jutten, C., and Hérault, J. (1991). Blind separation of of alert monkeys. Neuron 15: 843–856.
sources, part 1. An adaptive algorithm based on neuro-
Kaplan, C. A., and Simon, H. A. (1990). In search of in-
mimetic architecture. Signal Proc. 24: 1–10.
sight. Cogn. Psychol. 22: 374–419.
Kaas, J. H. (1991). Plasticity of sensory and motor maps in
Kapur, N., Abbott, P., Footitt, D., and Millar, J. (1996).
adult mammals. Annu. Rev. Neurosci. 14: 137–167.
Long-term perceptual priming in transient global amnesia.
Kaas, J. H. (1994). The reorganization of sensory and motor Brain Cogn. 31: 63–74.
maps in adult mammals. In The Cognitive Neurosciences, ed.
Karhunen, J., Oja, E., Wang, L., Vigario, R., and Joutse-
M. A. Gazzaniga. Cambridge, MA: MIT Press.
nalo, J. (1997). A class of neural networks for independent
Kaas, J. H. (1999). Is most of neural plasticity in the tha- component analysis. IEEE Trans. Neural Networks 8: 486–
lamus cortical? Proc. Natl. Acad. Sci. U S A 96: 7622–7623. 504.
Kaas, J., and Ebner, F. (1998). Intrathalamic connections: Karhunen, J., Wang, L., and Joutsensalo, J. (1995). Neural
A new way to modulate cortical plasticity? Nat. Neurosci. estimation of basis vectors in independent component
1: 341–342. analysis. In Proceedings of the International Conference on Neural
Networks, Paris.
Kaas, J. H., and Florence, S. L. (1997). Mechanisms of re-
organization in sensory systems of primates after peripheral Karni, A., and Bertini, G. (1997). Learning perceptual
nerve injury. Adv. Neurol. 73: 147–158. skills: Behavioral probes into adult cortical plasticity. Curr.
Opin. Neurobiol. 7: 530–535.
Kaas, J. H., Florence, S. L., and Jain, N. (1999). Subcortical
contributions to massive cortical reorganizations. Neuron Karni, A., and Sagi, D. (1991). Where practice makes perfect
22: 657–660. in texture discrimination: Evidence for primary visual cortex
plasticity. Proc. Natl. Acad. Sci. U S A 88: 4966–4970.
414 References
Karni, A., and Sagi, D. (1993). The time-course of learning matosensory cerebral cortex following digit amputation.
a visual skill. Nature 365: 250–252. Somatosens. Res. 2: 49–81.
Karni, A., and Sagi, D. (1995). A memory system in the Kemp, R., Pike, G., White, P., and Musselman, A. (1996).
adult visual cortex. In B. Julesz and I. Kovács, eds., Matura- Perception and recognition of normal and negative faces:
tional Windows and Adult Cortical Plasticity. SFI Studies in the The role of shape from shading and pigmentation cues.
Sciences of Complexity, vol. 24. Reading MA: Addison- Perception 25: 37–52.
Wesley.
Kemp, R., Towell, N., and Pike, G. (1997). When seeing
Karni, A., Meyer, G., Jazzard, P., Adams, M. M., Turner, should not be believing: Photographs, credit cards and
R., and Ungerleider, L. G. (1995). Functional MRI evi- fraud. Appl. Cogn. Psychol. 11: 211–222.
dence for adult motor plasticity during motor skill learn-
Kempermann, G., Kuhn, H. G., and Gage, F. H. (1997).
ing. Nature 377: 155–158.
More hippocampal neurons in adult mice living in an en-
Karni, A., Meyer, G., Rey-Ipolito, C., Jezzard, P., Adams, riched environment. Nature 386: 493–495.
M. M., Turner, R., and Ungerleider, L. G. (1998). The
Kerpelman, L. C. (1967). Stimulus dimensionality and
acquisition of skilled motor performance: Fast and slow
manipulability in visual perceptual learning. Child Dev. 38:
experience driven changes in primary motor cortex. Proc.
563–571.
Natl. Acad. Sci. U S A 96: 861–868.
Kersteen-Tucker, Z. (1991). Long-term repetition prim-
Karni, A., Tanne, D., Rubenstein, B. S., Askenasy, J. J.,
ing with symmetrical polygons and words. Mem. Cognit.
and Sagi, D. (1994). Dependence on REM sleep of over-
19: 37–43.
night improvement of a perceptual skill. Science 265: 679–
682. Kew, J. M., Ridding, M. C., Rothwell, J. C., Passingham,
R. E., Leigh, P. N., Sooriakumaran, D., Frackowiack,
Karni, A., Weisberg, J., Lalonde, F., and Ungerleider,
R. S. J., and Brooks, D. J. (1994). Reorganization of cor-
L. G. (1995). An fMRI study of human visual cortex plas-
tical blood flow and transcranial magnetic stimulation maps
ticity. Soc. Neurosci. Abstr. 21: 276.
in human subjects after upper limp amputation. J. Neuro-
Kasten E., and Sabel B. A. (1995). Visual field enlargement physiol. 72: 2517–2524.
after computer training in brain-damaged patients with
Kilgard, M. P., and Merzenich, M. M. (1995). Anticipated
homonymous deficits: An open pilot trial. Restor. Neurol.
stimuli across skin. Nature 373: 663.
Neurosci. 8: 113–127.
Kilgard, M., and Merzenich, M. M. (1998). Cortical map
Kasten, E., Wüst, S., Behrens-Baumann, W., and Sabel, B.
reorganization enabled by nucleus basalis activity. Science
A. (1998). Computer-based training for the treatment of
279: 1715–1718.
partial blindness. Nat. Med. 4: 1083–1087.
Killcross, A. S., Kiernan, M. J., Dwyer, D., and West-
Katz, L. C., and Callaway, E. M. (1992). Development
brook, R. F. (1998). E¤ects of retention interval on latent
of local circuits in mammalian visual cortex. Annu. Rev.
inhibition and perceptual learning. Q. J. Exp. Psychol. B.
Neurosci. 15: 31–56.
51: 59–74.
Katz, L. C., and Shatz, C. J. (1996). Synaptic activity and the
Kimura, A., Melis, F., and Asanuma, H. (1996). Long-
construction of cortical circuits. Science 274: 1133–1138.
lasting changes of neuronal activity in the motor cortex of
Katz, L. C., Burkhalter, A., and Dreyer, W. J. (1984). cats. NeuroReport 22: 869–872.
Fluorescent latex microspheres as a retrograde neuronal
Kimura, F., Nishigori, A., Shirokawa, T., and Tsumoto, T.
marker for in vivo and in vitro studies of visual cortex.
(1989). Long-term potentiation and n-methyl-d-aspartate
Nature 310: 498–500.
receptors in the visual cortex of young rats J. Physiol. 414:
Kelahan, A. M., and Doetsch, G. S. (1984). Time- 125–144.
dependent changes in the functional organization of so-
References 415
King, D. L., Shanks, S. C., and Hart, L. L. (1996). Dis- Kobatake, E., and Tanaka, K. (1998). E¤ects of shape-
crimination learning decreases perceived similarity accord- discrimination training on the selectivity of inferotemporal
ing to an objective measure. Psychol. Res. 59: 187–195. cells in adult monkeys. J. Neurophysiol. 80: 324–330.
Kirby, M., and Sirovich, L. (1990). Applications of the Kobatake, E., Wang, G., and Tanaka, K. (1994). Neuronal
Karhunen-Loeve procedure for the characterisation of hu- selectivities to complex object features in the ventral visual
man faces. IEEE Trans. Patt. Recog. Mach. Intell. 12: 103– pathway of the macaque cerebral cortex. J. Neurophysiol.
108. 71: 856–867.
Kirkwood, A., and Bear, M. F. (1994). Hebbian synapses Kobotake, E., Wang, G., and Tanaka, K. (1998). E¤ects of
in visual cortex. J. Neurosci. 14: 1634–1645. shape-discrimination training on the selectivity of infero-
temporal cells in adult monkeys. J. Neurophysiol. 80: 324–
Kirkwood, A., Rioult, M. G., and Bear, M. F. (1996).
330.
Experience-dependent modification of synaptic plasticity
in visual cortex. Nature 381: 526–528. Kodman, F. Jr. (1981). Perceptual-motor learning with
moderately retarded persons. Percept. Mot. Skills 53: 25–26.
Kisvárday, Z. F., and Eysel, U. T. (1992). Cellular orga-
nization of reciprocal patchy networks in layer III of cat Koenderink, J., and Doorn, A. (1991). A‰ne structure
visual cortex (area 17). Neuroscience 46: 275–286. from motion. J. Opt. Soc. Am. A 8: 377–385.
Kisvárday, Z. F., Cowey, A., Hodgson, A. J., and Somogyi, Ko¤ka, K. (1935). Principles of Gestalt Psychology. New
P. (1986). The relationship between GABA immunore- York: Harcourt Brace.
activity and labelling by local uptake of [ 3 H]GABA in the
Köhler, W. (1925). The Mentality of Apes. London: Rout-
striate cortex of monkey. Exp. Brain Res. 62: 89–98.
ledge and Kegan Paul.
Kisvárday, Z. F., Toth, E., Rausch, M., and Eysel, U. T.
Köhler, W. (1947). Gestalt Psychology. New York:
(1997). Orientation-specific relationship between popu-
Liveright.
lations of excitatory and inhibitory lateral connections in
the visual cortex of the cat. Cereb. Cortex 7: 605–618. Kohonen, T. (1982). Self-organized formation of topo-
logically correct feature maps. Biol. Cybern. 43: 59–69.
Kleim, A. J., Swain, R. A., Armstrong, K. A., Napper,
R. M. A., Jones, T. A., and Greenough, W. T. (1998). Kolb, B. (1999). Synaptic plasticity and the organization of
Selective synaptic plasticity within the cortex following behavior after early and late brain injury. Can. J. Exp.
complex motor skill learning. Neurobiol. Learn. Mem. 69: Psychol. 53: 62–75.
274–289. Kolb, B., and Wishaw, I. Q. (1998). Brain plasticity and
Knierim, J. J., and Van Essen, D. C. (1992). Neuronal behavior. Annu. Rev. Psychol. 49: 43–64.
responses to static texture patterns in area V1 of the alert König, P., Engel, A. K., Löwel, S., and Singer, W. (1993).
macaque monkey. J. Neurophysiol. 67: 961–980. Squint a¤ects synchronization of oscillatory responses in
Knierim, J. J., Kudrimoti, H. S., and McNaughton, B. L. cat visual cortex. Eur. J. Neurosci. 5: 501–508.
(1995). Place cells, head direction cells, and the learning of Konorski, J. (1967). Integrative Activity of the Brain. Chicago:
landmark stability. J. Neurosci. 15: 1648–1659. University of Chicago Press.
Knight, B., and Johnston, A. (1997). The role of move- Kopp, B., Kunkel, A., Mühlnickel, W., Villringer, K.,
ment in face recognition. Vis. Cognit. 4: 265–274. Taub, E., and Flor, H. (1999). Plasticity in motor system
Kobatake, E., and Tanaka, K. (1994). Neuronal selectiv- correlated with therapy-induced improvement of move-
ities to complex object features in the ventral visual path- ment after stroke. NeuroReport 10: 807–810.
way of the macaque cerebral cortex. J. Neurophysiol 71: Kornblith, H. (1985). Naturalizing Epistemology. Cam-
856–867. bridge, MA: MIT Press.
416 References
Kossel, A., Bonhoe¤er, T., and Bolz, J. (1990). Non- methiodide on receptive field properties of neurons in real
Hebbian synapses in rat visual cortex. NeuroReport 1: 115– and simulated whisker barrels. J. Neurophysiol. 75: 547–
118. 560.
Kovach, J. K. (1985). Constitutional biases in early per- LaBerge, D. (1976). Perceptual learning and attention. In
ceptual learning: III. Similarities and di¤erences between W. K. Estes, ed., Handbook of Learning and Cognitive Pro-
artificially selected and imprinted color preferences in quail cesses, vol. 4, pp. 237–273. Hillsdale, NJ: Erlbaum.
chicks (Coturnix coturnix japonica). J. Comp. Psychol. 99:
Lamdan, Y., Schwartz, J., and Wolfson, H. (1988). Object
35–46.
recognition by a‰ne invariant matching. In Proceedings of
Kovach, J. K., Fabricius, E., and Fält, L. (1966). Relation- the IEEE Conference on Computer Vision and Pattern Recog-
ships between imprinting and perceptual learning. J. Comp. nition, vol. 1, pp. 335–344.
Physiol. Psychol. 61: 449–454.
Land, E. H. (1983). Recent advances in Retinex theory
Krekling, S., Tellevik, J. M., and Nordvik, H. (1989). and some implications for cortical computations. Proc.
Tactual learning and cross-modal transfer of an oddity Natl. Acad. Sci. U S A 80: 5163–5169.
problem in young children. J. Exp. Child Psychol. 47: 88–
Land, E. H., and McCann, J. J. (1971). Lightness and
96.
Retinex theory. J. Opt. Soc. Am. 61: 1–11.
Krüger, J. (1989). Multiple recordings of neuronal prop-
Lander, K. (1999). The role of dynamic information in the
erties and spatial distributions in monkey visual cortex. J.
recognition of famous faces. Unpublished PhD thesis,
Physiol. 413.
University of Stirling.
Krupa, D. J., Ghazanfar, A. A., and Nicolelis, M. A. (1999).
Lander, K., and Bruce, V. (2000). Recognizing famous
Immediate thalamic sensory plasticity depends on cortico-
faces: Exploring the benefits of facial motion. Ecol. Psychol.
thalamic feedback. Proc. Natl. Acad. Sci. U S A 96: 8200–
12: 259–272.
8205.
Lander, K., Christie, F., and Bruce, V. (1999). The role of
Ku¿er, S. W. (1953). Discharge patterns and functional
movement in the recognition of famous faces. Mem. Cognit.
organization of the mammalian retina. J. Neurophysiol. 16:
27: 974–985.
37–68.
Landy, M., Maloney, L., Johnston, E., and Young, M.
Kujala, T., Alho, K., Huotilainen, M., Ilmoniemi, R. J.,
(1995). Measurement and modeling of depth cue combi-
Lehtokoki, A., Leinonen, A., Rinne, T., Salonen, O.,
nation: In defense of weak fusion. Vision Res. 35(3): 389–
Snikkonen, J., Standertskjöld-Nordenstam, C.-G., and
412.
Näätänen, R. (1997). Electrophysiological evidence for
cross-modal plasticity in humans with early- and late-onset Laubach, M., Wessberg, J., and Nicolelis, M. A. (2000).
blindness. Psychophysiology 34: 213–16. Cortical ensemble activity increasingly predicts behaviour
outcomes during learning of a motor task. Nature 405: 567–
Kujala, T., Alho, K., Kekoni, J., Hämäläinen, H., Reini-
571.
kainen, K., Salonen, O., Standertskjöld, C. G., and
Näätänen, R. (1995). Auditory and somatosensory event- Law, C. C., and Cooper, L. N. (1994). Formation of re-
related potentials in early blind humans. Exp. Brain Res. ceptive fields in realistic visual environments according to
104: 519–526. the Bienenstock, Cooper and Munro (BCM) theory. Proc.
Natl. Acad. Sci. U S A 91: 7797–7801.
Kulikowski, J. J., Abadi, R., and King-Smith, P. E. (1973).
Orientation selectivity of grating and line detectors in Lawrence, D. H. (1952). The transfer of discrimination
human vision. Vision Res. 13: 1479–1486. along a continuum. J. Comp. Physiol. Psychol. 45: 511–516.
Kyriazi, H. T., Carvell, G. E., Brumberg, J. C., and Simons, Layton, A. (1972). Body imagery in perceptual learning.
D. J. (1996). Quantitative e¤ects of GABA and bicuculline Am. J. Optom. Arch. Am. Acad. Optom. 49: 840–846.
References 417
Leder, H., and Bruce, V. (1998). Local and relational Levitt, L., and Bennett, T. L. (1975). The e¤ects of crowd-
aspects of facial distinctiveness. Q. J. Exp. Psychol. 51: 449– ing under di¤erent rearing conditions on emotionality and
473. transfer of perceptual learning. Behav. Biol. 15: 65–72.
Lee, C. J., and Whitsel, B. L. (1992). Mechanisms under- Levy, W. J., Amassian, V. E., Traad, M., and Cadwell, J.
lying somatosensory cortical dynamics: 1. In vivo studies. (1990). Focal magnetic coil stimulation reveals motor cor-
Cereb. Cortex 2: 81–106. tical system reorganized in humans after traumatic hemi-
plegia. Brain Res. 510: 130–134.
Lee, D. K., Koch, C., and Braun, J. (1997). Spatial vision
thresholds in the near absence of attention. Vision Res. 37: Lewicki, M. S., and Olshausen, B. A. (1999). Probabilistic
2409–2418. framework for the adaptation and comparison of image
codes. J. Opt. Soc. Am. A 16: 1587–1601.
Lee, T.-W., Bell, A. J., and Lambert, R. (1997). Blind
separation of delayed and convolved sources. In Advances Li, L., Miller, E. K., and Desimone, R. (1993). The rep-
in Neural Information Processing Systems, vol. 9, pp. 758– resentation of stimulus familiarity in anterior inferior tem-
764. Cambridge, MA: MIT Press. poral cortex. J. Neurophysiol. 69: 1918–1929.
Lee, T.-W., Girolami, M., and Sejnowski, T. J. (1999). Liberman, A. M., Harris, K. S., Ho¤man, H. S., and
Independent component analysis using an extended info- Gri‰th, B. C. (1957). The discrimination of speech sounds
max algorithm for mixed subgaussian and supergaussian within and across phoneme boundaries. J. Exp. Psychol. 54:
sources. Neural Comput. 11: 417–441. 358–368.
Leek, M. R., and Watson, C. S. (1988). Auditory percep- Liepert, J., Miltner, W. H. R., Bauder, H., Sommer, M.,
tual learning of tonal patterns. Percept. Psychophys. 43: 389– Dettmers, C., Taub, E., and Weiller, C. (1998). Motor
394. cortex plasticity during contraint-induced movement ther-
apy in stroke patients. Neurosci. Lett. 250: 5–8.
Leen, T. K., and Kambhatla, N. (1994). Fast non-linear
dimension reduction. In J. D. Cowan, G. Tesauro, and J. Liepert, J., Tegentho¤, M., and Malin, J. P. (1995).
Alspector, eds., Advances in Neural Information-Processing Sys- Changes of cortical motor area size during immobilization.
tems, vol. 6, pp. 152–159. San Francisco: Morgan Kaufmann. Electroenceph. Clin. Neurophysiol. 97: 382–386.
Legge, G. E., and Foley, J. M. (1980). Contrast masking in Liepert, J., Terborg, C., and Weiller, C. (1999). Motor
human vision. J. Opt. Soc. Am. 70: 1458–1471. plasticity induced by synchronized thumb and foot move-
ments. Exp. Brain Res. 125: 435–439.
Leonards, U., Rettenbach, R., and Sireteanu, R. (1998).
Parallel visual search is not always e¤ortless. Brain Res. Linsker, R. (1988). Self-organization in a perceptual net-
Cogn. Brain Res. 7: 207–213. work. Computer 21: 105–117.
Lepore, F., and Guillemot, J. P. (1982). Visual receptive Linsker, R. (1992). Local synaptic learning rules su‰ce to
field properties of cells innervated through the corpus cal- maximize mutual information in a linear network. Neural
losum in the cat. Exp. Brain Res. 46: 413–424. Comput. 4: 691–702.
Levänen, S., Jousmäki, V., and Hari, R. (1998). Vibration- Lissauer, I. (1890). Ein Fall von Seelenblindheit nebst
induced auditory cortex activation in a congentially deaf einem Beitrage zur Theorie derselben. Archiv für Psychiatrie
adult. Curr. Biol. 8: 869–872. und Nervenkrankheiten 21: 222–270.
Levi, D. M., and Polat, U. (1996). Neural plasticity in Liu, Z. (1999). Perceptual learning in motion discrimina-
adults with amblyopia. Proc. Natl. Acad. Sci. U S A 93: tion that generalizes across motion directions. Proc. Natl.
6830–6834. Acad. Sci. U S A 96: 14085–14087.
Levin, D. T., and Beale, J. M. (2000). Categorical percep- Liu, Z., and Vaina, L. M. (1998). Simultaneous learning of
tion occurs in newly learned faces, other-race faces, and motion discrimination in two directions. Cogn. Brain Res.
inverted faces. Percept. Psychophys. 62(2): 386–401. 6: 347–349.
418 References
Liu, Z., and Weinshall, D. (2000). Mechanisms of gener- Löwel, S., and Singer, W. (1992). Selection of intrinsic
alization in perceptual learning. Vision Res. 40: 97–109. horizontal connections in the visual cortex by correlated
neuronal activity. Science 255: 209–212.
Livingstone, M. S., and Hubel, D. H. (1984). Specificity of
intrinsic connections in primate primary visual cortex. J. Lübke, J., and Albus, K. (1992). Rapid rearrangement of
Neurosci. 4: 2830–2835. intrinsic tangential connections in the striate cortex of
normal and dark-reared kittens: Lack of exuberance be-
Lobley, K., and Walsh, V. (1998). Perceptual learning in
yond the second postnatal week. J. Comp. Neurol. 323: 42–
visual conjunction search. Perception 27: 1245–1255.
58.
Locke, J. (1690/1939). An Essay Concerning Human Under-
Ludvigh, E. (1953). Direction sense of the eye. Am. J.
standing. In The English Philosphers from Bacon to Mill, ed.
Ophthalmol. 36: 139–142.
E. A. Burtt. New York: Random House.
Lueschow, A., Miller, E. K., and Desimone, R. (1994).
Locke, J. (1708). Some Familiar Letters between Mr. Locke and
Inferior temporal mechanisms for invariant object recog-
Several of His Friends. London: A. and J. Churchill.
nition. Cereb. Cortex 5: 523–531.
Locke, J. (1721). An Essay Concerning Human Understand-
Luhmann, H. J., and Prince, D. A. (1991). Control of
ing, 8th ed. London: A. Churchill and A. Manship; and
NMDA receptor–mediated activity by GABAergic mech-
Sold by W. Taylor, at the Ship and Black Swan in Pater-
anisms in mature and developing rat neocortex. Dev. Brain
noster Row.
Res. 54: 287–290.
Lockwood, A. H., Salvi, R. J., Coad, M. L., Towsley,
Luhmann, H. J., Martı́nez-Millán, L., and Singer, W.
M. L., Wack, D. S., and Murphy, B. W. (1998). The func-
(1986). Development of horizontal intrinsic connections in
tional neuroanatomy of tinnitus: Evidence for limbic sys-
cat striate cortex. Exp. Brain Res. 63: 443–448.
tem links and neural plasticity. Neurology 50: 114–120.
Luhmann, H. J., Singer, W., and Martı́nez-Millán, L.
Logothetis, N. K., and Pauls, J. (1995). Psychophysical and
(1990). Horizontal interactions in cat striate cortex: 1.
physiological evidence for viewer-centered object repre-
Anatomical substrate and postnatal development. Eur. J.
sentations in the primate. Cereb. Cortex 5: 270–288.
Neurosci. 2: 344–357.
Logothetis, N. K., Pauls, J., and Poggio, T. (1995). Shape
Lund, J. S. (1973). Organization of neurons in the visual
representation in the inferior temporal cortex of monkeys.
cortex, area 17, of the monkey (Macaca mulatta). J. Comp.
Curr. Biol. 5: 552–563.
Neurol. 147: 455–496.
Logothetis, N. K., and Sheinberg, D. L. (1996). Visual
Lund, J. S. (1988). Anatomical organization of macaque
object recognition. Annu. Rev. Neurosci. 19: 577–621.
monkey striate visual cortex. Annu. Rev. Neurosci. 11: 253–
Losada, M. A., and Mullen, K. T. (1994). The spatial tun- 288.
ing of chromatic mechanisms identified by simultaneous
Lund, R. D., Mitchell, D. E., and Henry, G. H. (1978).
masking. Vision Res. 34: 331–341.
Squint-induced modification of callosal connections in
Lotze, M., Grodd, W., Birbaumer, N., Erb, M., Huse, E., cats. Brain Res. 144: 169–172.
and Flor, H. (1999). Does use of a myoelectric prosthesis
Luria, A. R. (1963). The Working Brain. Harmondsworth:
prevent cortical reorganization and phantom limb pain?
Penguin.
Nat. Neurosci. 2: 501–502.
Lütkenhöner, B. (1996). Current dipole localization with
Lowe, D. (1984). Perceptual organization and visual rec-
an ideal magnetometer system. IEEE 43(11): 1049–1061.
ognition. Ph.D. diss., Stanford University.
Lütkenhöner, B., Hoke, M., and Pantev, C. (1990). Use of
Löwel, S. (1994). Ocular dominance column development:
biomagnetic examination procedures in audiology. Biomed
Strabismus changes the spacing of adjacent columns in cat
Tech 35(Suppl. 3): 154–155.
visual cortex. J. Neurosci. 14: 7451–7468.
References 419
Lutzer, V. D. (1986). Perceptual learning by educable lesions on learning and memory in monkeys. Behav. Neu-
mentally retarded, average, and gifted children of primary rosci. 109: 212–226.
school age. Percept. Mot. Skills 62: 959–966.
Marill, T. (1991). Emulating the human interpretation of
Lutzer, V. D. (1987). Perceptual learning of a non-focal line-drawings as three-dimensional objects. Intl. J. Comput.
color discrimination between ages three and six. J. Gen. Vis. 6: 147–161.
Psychol. 114: 273–279.
Markam, H., Lübke, J., Frotscher, M., and Sakmann, B.
Mackintosh, N. J., Kaye, H., and Bennett, C. H. (1991). (1997). Regulation of synaptic e‰cacy by coincidence of
Perceptual learning in flavour aversion conditioning. Q. J. postsynaptic APs and EPSPs. Science 275: 213–15.
Exp. Psychol. B. 43: 297–322.
Markowitsch, H. J., and Harting, C. (1996). Interdepen-
Maclin, E. L., Rose, D. F., Knight, J. E., Orrison, W. W., dence of priming performance and brain-damage. Int. J.
and Davis, L. E. (1994). Somatosensory evoked magnetic Neurosci. 85: 291–300.
fields in patients with stroke. Electroenc. Clin. Neurophysiol.
Marks, L. E., Galanter, E., and Baird, J. C. (1995). Binaural
91: 468–475.
summation after learning psychophysical functions for
Maddox, W. T., and Bohil, C. J. (2001). Feedback e¤ects loudness. Percept. Psychophys. 57: 1209–1216.
on cost-benefit learning in perceptual categorization. Mem.
Markson, L., and Bloom, P. (1997). Evidence against a
Cognit. 29: 598–615.
dedicated system for word learning in children. Nature 385:
Ma¤ei, L., Fiorentini, A., and Bisti, S. (1973). Neural 813–815.
correlate to perceptual adaptation to gratings. Science 182:
Marquet, J., Van Durme, M., Lammens, J., Collier, R.,
1036–1103.
Peeters, S., and Bosiers, W. (1986). Acoustic simulation
Makeig, S. (1993). Auditory event-related dynamics of the experiments with preprocessed speech for an 8-channel
EEG spectrum and e¤ects of exposure to tones. Electroenc. cochlear implant. Audiology 25: 353–362.
Clin. Neurophysiol. 86: 283–293.
Marr, D. (1970). A theory for cerebral neocortex. Proc. R.
Malach, R., Amir, Y., Harel, M., and Grinvald, A. (1993). Soc. Lond. B Biol. Sci. 176: 161–234.
Relationship between intrinsic connections and functional
Marr, D. (1971). Simple memory: A theory for archicortex.
architecture revealed by optical imaging and in vivo tar-
Philos. Trans. R. Soc. Lond. B Biol. Sci. 262: 23–81.
geted biocytin injections in primate striate cortex. Proc.
Natl. Acad. Sci. U S A 90: 10469–10473. Marr, D. (1982). Vision. San Francisco: Freeman.
Malach, R., Tootell, R. B. H., and Malonek, D. (1994). Marr, D., and Hildreth, E. (1980). Theory of edge-
Relationship between orientation domains, cytochrome detection. Proc. R. Soc. Lond. B Biol. Sci. 207: 187–217.
oxidase stripes, and intrinsic horizontal connections in Marr, D., and Nishihara, H. (1978). Representation and
squirrel monkey area V2. Cereb. Cortex 4: 151–165. recognition of the spatial organization of three-dimensional
Maldonado, P. E., and Gerstein, G. L. (1996a). Reorgani- structure. Proc. R. Soc. Lond. B Biol. Sci. 200: 269–294.
zation in the auditory cortex of the rat induced by intra- Marr, D., and Poggio, T. (1977). From understanding
cortical microstimulation: A multiple single-unit study. computation to understanding neural circuitry. Neurosci.
Exp. Brain Res. 112: 420–430. Res. Prog. Bull. 15: 470–488.
Maldonado, P. E., and Gerstein, G. L. (1996b). Neuronal Martin, K. A. C., and Whitteridge, D. (1984). Form,
assembly dynamics in the rat auditory cortex during reor- function and intracortical projections of spiny neurones in
ganization induced by intracortical microstimulation. Exp. the striate visual cortex of the cat. J. Physiol. 253: 463–504.
Brain Res. 112: 431–441.
Martini, P., Girard, P., Morrone, M. C., and Burr, D.
Malkova, L., Mishkin, M., and Bachevalier J. (1995). (1996). Sensitivity to spatial phase at equiluminance. Vision
Long-term e¤ects of selective neonatal temporal lobe Res. 36: 1153–1162.
420 References
Masson, M. E. (1986). Identification of typographically Melamed, L. E., and Arnett, W. B. (1984). The e¤ect of
transformed words: Instance-based skill acquisition. J. Exp. familial sinistrality on perceptual learning. Neuropsychologia
Psychol. Learn. Mem. Cogn. 12: 479–488. 22: 495–502.
Mato, G., and Sompolinsky, H. (1996). Neural network Merigan, W. H., Nealy, T. A., and Maunsell, J. H. R.
models of perceptual learning of angle discrimination. (1993). Visual e¤ects of lesions of cortical area V2 in
Neural Comput. 8: 270–299. macaques. J. Neurosci. 13: 3180–3191.
Matthews, N., Liu, Z., Geesaman, B. J., and Qian, N. Merzenich, M. M., and Jenkins, W. M. (1993). Reorga-
(1999). Perceptual learning on orientation and direction nization of cortical representations of the hand following
discrimination. Vision Res. 39: 3692–3701. alterations of skin inputs induced by nerve injury, skin
island transfer and experience. J. Hand Ther. 6: 89–103.
Mayer, R. E. (1995). The search for insight: Grappling
with Gestalt psychology’s unanswered questions. In The Merzenich, M. M., and White, M. (1980). Coding con-
Nature of Insight, ed. R. J. Sternberg and J. E. Davidson. siderations in design of cochlear prostheses. Ann. Otol.
Cambridge, MA: MIT Press. Rhinol. Laryngol. 89: 84–87.
McClelland, J. L., and Rumelhart, D. E. (1981). An Merzenich, M. M., Jenkins, W. M., Johnston, P.,
interactive activation model of context e¤ects in letter Schreiner, C., Miller, S. L., and Tallal, P. (1996). Temporal
perception: 1. An account of basic findings. Psychol. Rev. processing deficits of language-learning impaired children
88: 375–407. ameliorated by training. Nature 271: 77–80.
McGuire, B. A., Gilbert, C. D., Rivlin, P. K., and Wiesel, Merzenich, M. M., Kaas, J. H., Wall, J., Nelson, R. J., Sur,
T. N. (1991). Targets of horizontal connections in macaque M., and Felleman, D. (1983a). Topographic reorganization
primary visual cortex. J. Comp. Neurol. 305: 370–392. of somatosensory cortical areas 3b and 1 in adult monkeys
following restricted dea¤erentation. Neuroscience 8: 33–55.
McIntosh, A. R., Rajah, M. N., and Lobaugh, N. J.
(1999). Interactions of prefrontal cortex in relation to Merzenich, M. M., Kaas, J. H., Wall, J., Sur, M., Nelson,
awareness in sensory learning. Science 284: 1531–1533. R. J., and Felleman, D. (1983b). Progression of changes
following median nerve section in the cortical representa-
McKee, S. P., and Westheimer, G. (1978). Improvement in
tion of the hand in areas 3b and 1 in adult owl and squirrel
vernier acuity with practice. Percept. Psychophys. 24: 258–
monkeys. Neuroscience 10: 639–665.
262.
Merzenich, M. M., Nelson, R. J., Stryker, M. P., Cynader,
McLaren, I. P. (1997). Categorization and perceptual
M. S., Schoppmann, A., and Zook, J. M. (1984). Somato-
learning: An analogue of the face inversion e¤ect. Q. J.
sensory cortical map changes following digit amputation in
Exp. Psychol. A. 50: 257–273.
adult monkeys. J. Comp. Neurol. 224: 591–605.
Medin, D. L. (1989). Concepts and conceptual structure.
Merzenich, M. M., Recanzone, G., Jenkins, W. M., Allard,
Am. Psychol. 44: 1469–1481.
T. T., and Nudo, R. J. (1988). Cortical representational
Meegan, D. V., Aslin, R. N., and Jacobs, R. A. (2000). plasticity. In P. Rakic and W. Singer eds., Neurobiology of
Motor timing learned without motor training. Nat. Neu- Neocortex, 41–67. New York: Wiley.
rosci. 3: 860–862.
Merzenich, M. M., Schreiner, C., Jenkins, W., and Wang,
Meinhardt, G. (2001). Learning a grating discrimination X. (1993). Neural mechanisms underlying temporal inte-
task broadens human spatial frequency tuning. Biol. Cybern. gration, segmentation, and input sequence representation:
86: 383–400. Some implications for the origin of learning disabilities.
Meinhardt, G., and Grabbe, Y. (in press). Attentional Ann. N. Y. Acad. Sci. 682: 1–22.
control in learning to discriminate bars and gratings. Exp. Merzenich, M., Wright, B., Jenkins, W., Xerri, C., Byl,
Brain Res. N., Miller, S., and Tallal, P. (1996). Cortical plasticity
References 421
underlying perceptual, motor, and cognitive skill devel- Miyashita, Y., Date, A., and Okuno, H. (1993). Config-
opment: implications for neurorehabilitation. Cold Spring urational encoding of complex visual forms by single neu-
Harb. Symp. Quant. Biol. 61: 1–8. rons of monkey temporal cortex. Neuropsychologia 31:
1119–1131.
Miller, E. K., Li, L., and Desimone, R. (1991). A neural
mechanism for working and recognition memory in infe- Miyashita, Y., Okuno, H., Tokuyama, W., Ihara, T., and
rior temporal cortex. Science 254: 1377–1379. Nakajima, K. (1996). Feedback signal from medial tem-
poral lobe mediates visual associative mnemonic codes of
Miller, K. D. (1988). Correlation-based models of neural
inferotemporal neurons. Cogn. Brain Res. 5: 81–86.
development. In M. Gluck and D. Rumelhart, eds., Neu-
roscience and Connectionist Theory, pp. 267–353. Hillsdale, Mogilner, A., Grossman, J. A. I., Ribary, U., Joliot, M.,
NJ: Erlbaum. Volkmann, J., Rapaport, D., Beasley, R. W., and Llinás,
R. R. (1993). Somatosensory cortical plasticity in adult
Miller, K. D., Keller, J. B., and Stryker, M. P. (1989). Oc-
humans revealed by magnetoencephalography. Proc. Natl.
ular dominance column development: Analysis and simu-
Acad. Sci. U.S.A. 90: 3593–3597.
lation. Science 245: 605–615.
Mollon, J. D., and Danilova, M. V. (1996). Three remarks
Millodot, M. (1965). Stabilized retinal images and disap-
on perceptual learning. Spat. Vis. 10: 51–58.
pearance time. Br. J. Physiol. Opt. 22: 148–152.
Montague, P. R., and Sejnowski, T. J. (1994). The pre-
Miltner, W. H., Braun, C., Arnold, M., Witte, H., and
dictive brain: Temporal coincidence and temporal order in
Taub, E. (1999). Coherence of gamma-band EEG activity
synaptic learning mechanisms. Learn. Mem. 1: 1–33.
as a basis for associative learning. Nature 397: 434–436.
Moore, C., and Cavanagh, P. (1998). Recovery of 3D
Minsky, M., and Papert, S. (1969). Perceptrons. Cambridge,
volume from 2-tone images of novel objects. Cognition 67:
MA: MIT Press.
45–71.
Missal, M., Vogels, R., and Orban, G. A. (1997). Responses
Moore, D. R., and Kowalchuk, N. E. (1988). Auditory
of macaque inferior temporal neurons to overlapping
brainstem of the ferret: E¤ects of unilateral cochlear lesions
shapes. Cereb. Cortex 7: 758–767.
on cochlear nucleus volume, and projections to the infe-
Mitchison, G., and Crick, F. (1982). Long axons within the rior colliculus. J. Comp. Neurol. 272: 503–515.
striate cortex: Their distribution, orientation, and patterns
Moran, J., and Desimone, R. (1985). Selective attention
of connection. Proc. Natl. Acad. Sci. U S A 79: 3661–3665.
gates visual processing in extrastriate cortex. Science 229:
Mittmann, T., and Eysel, U. T. (2001). Increased synaptic 782–784.
plasticity in the surround of rat visual cortex lesions. Neuro-
Morgan, M. J., and Baldassi, S. (1997). How the human
Report 12: 3341–3347.
visual system encodes the orientation of a texture, and why
Mittmann, T., Luhmann, H. J., Schmidt-Kastner, R., it makes mistakes. Curr. Biol. 7: 999–1002.
Eysel, U. T., and Heinemann, U. (1994). Lesion-induced
Morrone, M. C., Burr, D. C., and Ma¤ei, L. (1982).
transient suppression of inhibitory function in rat neo-
Functional implications of cross-orientation inhibition of
cortex in vitro. Neuroscience 60: 891–906.
cortical visual cells: 1. Neurophysiological evidence. Proc.
Miyashita, Y. (1988). Neural correlate of visual associative R. Soc. Lond. B Biol. Sci. 216: 335–354.
long-term memory in the primate temporal cortex. Nature
Morrone, M. C., Burr, D. C., and Spinelli, D. (1989). Dis-
335: 817–820.
crimination of spatial phase in central and peripheral vision.
Miyashita, Y. (1993). Inferior temporal cortex: Where Vision Res. 29: 433–445.
visual perception meets memory. Annu. Rev. Neurosci. 16:
Moses, Y., Schechtman, G., and Ullman, S. (1990). Self-
245–263.
calibrated collinearity detector. Biol. Cybern. 63: 463–475.
422 References
Motter, B. C. (1993). Focal attention produces spatially Nachmias, J., and Sansbury, R. V. (1974). Grating con-
selective processing in visual cortical areas V1, V2, and V4 trast: Discrimination may be better than detection. Vision
in the presence of competing stimuli. J. Neurophysiol. 70: Res. 14: 1039–1042.
909–919.
Nachmias, J., and Weber, A. (1975). Discrimination of
Mountcastle, V. B., Steinmetz, M. A., and Romo, R. simple and complex gratings. Vision Res. 15: 217–223.
(1990). Frequency discrimination in the sense of flutter:
Nadal, J.-P., and Parga, N. (1994). Non-linear neurons in
Psychophysical measurements correlated with postcentral
the low noise limit: A factorial code maximizes informa-
events in behaving monkeys. J. Neurosci. 10: 3032–3044.
tion transfer. Network 5: 565–581.
Movshon, J. A., and Blakemore, C. B. (1973). Orientation
Nadol, J. B., Young, Y., and Glynn, R. J. (1989). Survival
specificity and spatial selectivity in human vision. Perception
of spiral ganglion cells in profound sensory neural hearing
2: 53–60.
loss: Implications for cochlear implantation. Ann. Otol.
Movshon, J. A., and Lennie, P. (1979). Pattern selective Rhinol. Laryngol. 98: 411–416.
adaptation in visual cortical neurones. Nature 278: 850–
Nagarajan, S. S., Blake, D. T., Wright, B. A., Byl, N., and
852.
Merzenich, M. M. (1998). Practice-related improvements
Mühlnickel, W., Elbert, T., Taub, E., and Flor, H. (1998). in somatosensory interval discrimination are temporally
Reorganization of auditory cortex in tinnitus. Proc. Natl. specific but generalize across skin location, hemisphere,
Acad. Sci. U.S.A. 95: 10340–10343. and modality. J. Neurosci. 18: 1559–1570.
Müller, M. M., Bosch, J., Elbert, T., Kreiter, A., Valdes Nakamura, H., Gattass, R., Desimone, R., and Unger-
Sosa, M., Valdes Sosa, P., Rockstroh, B. (1996). Visually leider, L. G. (1993). The modular organization of projec-
induced gamma band responses in human EEG. A link to tions from areas V1 and V2 to areas V4 and TEO in
animal studies. Exp. Brain Res. 112: 96–112. macaques. J. Neurosci. 13: 3681–3691.
Müller, M. M., Junghöfer, M., Elbert, T., and Rockstroh, Nakayama, K. (1991). The iconic bottleneck and the ten-
B. (1997). Visually induced gamma-band responses to co- uous link between early visual processing and perception.
herent and incoherent motion: A replication. NeuroReport In C. Blakemore, ed., Vision: Coding and E‰ciency, 411–
8(11): 2575–2579. 422. Cambridge: Cambridge University Press.
Mullen, K. T. (1985). The contrast sensitivity of human Nazir, T. A., and O’Regan, J. K. (1990). Some results on
colour vision to red-green and blue-yellow gratings. J. translation invariance in the human visual system. Spat.
Physiol. (Lond.) 359: 381–400. Vis. 5: 81–100.
Mumford, D. (1992). On the computational architecture Nelson, C. A. (1999). Neural plasticity and human devel-
of the neocortex: 2. The role of cortico-cortical loops. opment: The role of early experience in sculpting memory
Biol. Cybern. 66: 241–251. systems. Psychol. Sci. 8: 42–45.
Mundy, J., and Zisserman, A. (1992). Introduction: Nelson, C. A., and Bloom. F. (1997). Child development
Towards a new framework for vision. In J. Mundy and A. and neuroscience. Child Dev. 68(5): 970–987.
Zisserman, eds., Geometric Invariance in Computer Vision, pp.
Nelson, J. I., and Frost, B. J. (1978). Orientation-selective
1–39. Cambridge, MA: MIT Press.
inhibition from beyond the classical receptive field. Exp.
Murayama, Y., Fujita, I., and Kato, M. (1997). Contrasting Brain Res. 139: 359–365.
forms of synaptic plasticity in monkey inferotemporal and
Nelson, R. B., Friedman, D. P., O’Neill, J. B., Mishkin,
primary visual cortices. NeuroReport 8: 1503–1508.
M., and Routtenberg, A. (1987). Gratients of protein
Murray, E. A., Ga¤an, D., and Mishkin, M. (1993). Neural kinase C substrate phosphorylation in primate visual system
substrates of visual stimulus-stimulus association in rhesus peak in visual memory storage areas. Brain Res. 416: 387–
monkeys. J. Neurosci. 13: 4549–4561. 392.
References 423
Nevatia, R., and Binford, T. (1977). Description and rec- Nothdurft, H. C. (1992). Feature analysis and the role of
ognition of curved objects. Artif. Intell. 8: 77–98. similarity in preattentive vision. Percept. Psychophys. 52:
355–375.
Newell, F., Chiroro, P., and Valentine, T. (1999). Rec-
ognising unfamiliar faces: The e¤ects of distinctiveness and Nudo, R. J., Jenkins, W. M., and Merzenich, M. M.
view. Q. J. Exp. Psychol. A 52: 509–534. (1990). Repetitive microstimulation alters the cortical rep-
resentation of movements in adult rats. Somatosen. Mot. Res.
Nicolelis, M. A. L., ed. (1999). Methods in Neural Ensemble
7: 463–483.
Recordings. New York: CRC Press.
Nygaard, L. C., and Pisoni, D. B. (1998). Talker-specific
Nicolelis, M. A., Ghazanfar, A. A., Faggin, B. M., Votaw
learning in speech perception. Percept. Psychophys. 60: 355–
S., and Oliveira L. M. (1989). Reconstructing the engram:
376.
Simultaneous, multisite, many single neuron recordings.
Neuron 18: 529–537. Obata, S., Obata, J., Das, A., and Gilbert, C. D. (1999).
Molecular correlates of topographic reorganization in pri-
Nicolelis, M. A., Ghazanfar, A. A., Stambaugh, C. R.,
mary visual cortex following retinal lesions. Cereb. Cortex
Oliveira, L. M., Laubach, M., Chapin, J. K., Nelson, R. J.,
9: 238–248.
and Kaas, J. H. (1998). Simultaneous encoding of tactile
information by three primate cortical areas. Nat. Neurosci. Odom, R. D., McIntyre, C. W., and Neale, G. S. (1971).
1: 621–630. The influence of cognitive style on perceptual learning.
Child Dev. 42: 883–891.
Nicolelis, M. A., Katz, D., and Krupa, D. J. (1998). Po-
tential circuit mechanisms underlying concurrent thalamic O’Donnell, C., and Bruce, V. (2001). Familiarisation with
and cortical plasticity. Rev. Neurosci. 9: 213–224. faces selectively enhances sensitivity to changes made to
the eyes. Perception 30: 755–764.
Nicolelis, M. A., Lin, R. C., and Chapin, J. K. (1997).
Neonatal whisker removal reduces the discrimination of Ohl, F., and Scheich, H. (1997). Learning-induced dy-
tactile stimuli by thalamic ensembles in adult rats. J. Neu- namic receptive field changes in primary auditory cortex
rophysiol. 78: 1691–1706. (A1) of the unanaestetized mongolian gerbil. J. Comp.
Physiol. A181: 685–696.
Nicolelis, M. A., Lin, R. C., Woodward, D. J., and Chapin,
J. K. (1993). Induction of immediate spatiotemporal Ohzawa, I., Schlar, G., and Freeman, R. D. (1982). Con-
changes in thalamic networks by peripheral block of trast gain control in the cat’s visual cortex. Nature 298:
ascending cutaneous information. Nature 361: 533–536. 5871–5873.
Niebur, E., and Koch, C. (1994). A model for the neu- Oja, E. (1989). Neural networks, principal compo-
ronal implementation of selective visual attention based on nents and linear neural networks. Neural Networks 5: 927–
temporal correlation among neurons. J. Comput. Neurosci. 935.
1: 141–158.
O’Keefe, J., and Nadel, L. (1978). The Hippocampus as a
Nordern, K. W., Killackoy, H. P., and Kitzes, L. M. (1983). Cognitive Map. Oxford: Clarendon Press.
Ascending projections to the inferior colliculus following
Olshausen, B. A., and Field, D. J. (1996). Natural image
unilateral cochlear ablation in the neonatal gerbil, Mexiones
statistics and e‰cient coding. Network Comput. Neural Syst.
unguiculatus. J. Comp. Neurol. 214: 144–153.
7: 333–339.
Nosofsky, R. M. (1988). Exemplar-based accounts of rela-
Olshausen, B. A., and Field, D. J. (1997). Sparse coding
tions between classification, recognition, and typicality. J.
with an overcomplete basis set: a strategy employed by V1?
Exp. Psychol. Learn. Mem. Cognit. 14: 700–708.
Vision Res. 37: 3311–3325.
Nothdurft, H. C. (1985). Orientation sensitivity and tex-
Omohundro, S. M. (1987). E‰cient algorithms with neu-
ture segmentation in patterns with di¤erent line orienta-
ral network behavior. Complex Syst. 1: 273–347.
tion. Vision Res. 25(4): 551–560.
424 References
Oram, M. W., and Perrett, D. I. (1992). Time course of Pascual-Leone, A., Wassermann, E. M., Sadato, N., and
neural responses discriminating di¤erent views of the face Hallett, M. (1995). The role of reading activity on the
and head. J. Neurophysiol. 68: 70–84. modulation of motor cortical outputs to the reading hand
in Braille readers. Ann. Neurol. 38: 910–915.
O Scalaidhe, S. P., Wilson, F. A. W., and Goldman-Rakic,
P. S. (1997). Areal segregation of face-processing neurons Pashler, H. E. (1998). The Psychology of Attention. Cam-
in prefrontal cortex. Science 278: 1135–1138. bridge, MA: MIT Press.
Osgood, C. E. (1949). The similarity paradox in human Pavlov, I. P. (1927). Conditioned Reflexes, pp. 121–122.
learning: A resolution. Psychol. Rev. 56: 132–143. Oxford: Oxford University Press.
O’Toole, A. J., and Kersten, D. J. (1992). Learning to see Pei, X., Vidyasagar, T. R., Volgushev, M., and Creutzfeldt,
random-dot stereograms. Perception 21: 227–243. O. D. (1994). Receptive field analysis and orientation
selectivity of postsynaptic potentials of simple cells in cat
Otsu, Y., Kimura, F., and Tsumoto, T. (1995). Hebbian
visual cortex. J. Neurosci. 14: 7130–7140.
induction of LTP in visual cortex: Perforated patch-clamp
study in cultured neurons. J. Neurophysiol. 74: 2437– Pellegrino, L. J., Pellegrino, A. S., and Cushman, A. J.
2444. (1986). A Stereotaxic Atlas of the Rat Brain, 2nd ed. New
York: Plenum.
Owen, D. H., and Machamer, P. K. (1979). Bias-free im-
provement in wine discrimination. Perception 8: 199–209. Pelli, D. G. (1985). Uncertainty explains many aspects of
visual contrast detection and discrimination. J. Opt. Soc.
Palmer, S., Rosch, E., and Chase, P. (1981). Canonical
Am. A 2: 1508–1532.
perspective and the perception of objects. In J. Long and
A. Baddeley, eds., Attention and Performance IX, 131–151. Pentland, A. (1986). Perceptual organization and the rep-
Hillsdale, NJ: Erlbaum. resentation of natural form. Artif. Intell. 28: 293–331.
Pantev, C., Oostenveld, R., Engelien, A., Ross, B., Peres, R., and Hochstein, S. (1994). Modeling perceptual
Roberts, L. E., and Hoke, M. (1998). Increased auditory learning with multiple interacting elements: A neural net-
cortical representation in musicians. Nature 392: 811–814. work model describing early visual perceptual learning. J.
Comput. Neurosci. 1: 323–338.
Papathomas, T. V., Gorea, A., Feher, A., and Conway,
T. E. (1999). Attention-based texture segregation. Percept. Pernberg, J., Jirmann, K. U., and Eysel, U. T. (1998).
Psychophys. 61: 1399–1410. Structure and dynamics of receptive fields in the visual
cortex of the cat (area 18) and the influence of GABAergic
Parker, A. J., and Newsome, W. T. (1998). Sense and the
inhibition. Europ. J. Neurosci. 10: 3596–3606.
single neuron: Probing the physiology of perception.
Annu. Rev. Neurosci. 21: 227–277. Perrett, D., Hietanen, J., Oram, M., and Benson, P. (1992).
Organisation and functions of cells responsive to faces in
Pascual-Leone, A., Cammarota, A., Wassermann, E. M.,
the temporal cortex. Philos. Trans. R. Soc. Lond. B Biol.
Brasil-Neto, J. P., Cohen, L. G., and Hallett, M. (1993).
Sci. 335: 23–30.
Modulation of motor cortical outputs to the reading hand
of braille readers. Ann. Neurol. 34: 33–37. Perrett, D. I., Mistlin, A. J., and Chitty, A. J. (1987). Visual
neurones responsive to faces. Trends Neurosci. 10: 358–364.
Pascual-Leone, A., Peris, M., Tormos, J. M., and Catalá,
M. D. (1996). Reorganization of human cortical output Perrett, D. I., Rolls, E. T., and Caan, W. (1979). Temporal
maps following traumatic forearm amputation. Neuro- lobe cells of the monkey with visual responses selective for
Report 13(2): 2068–2070. faces. Neurosci. Lett. Suppl. S3: S358.
Pascual-Leone, A., and Torres, F. (1993). Plasticity of the Perrett, D. I., Rolls, E. T., and Caan, W. (1982). Visual
sensorimotor cortex representation of the reading finger in neurones responsive to faces in the monkey temporal cor-
Braille readers. Brain 116: 39–52. tex. Exp. Brain Res. 47: 329–342.
References 425
Perrett, D. I., Smith, P. A. J., Potter, D. D., Mistlin, A. J., Pigarev, I. N., Rizzolatti, G., and Scandolara, C. (1979).
Head, A. S., Milner, A. D., and Jeeves, M. A. (1984). Neurons responding to visual stimuli in the frontal lobe of
Neurones responsive to faces in the temporal cortex: macaque monkeys. Neurosci. Lett. 12: 207–212.
Studies of functional organization, sensitivity to identity
Pike, G. E., Kemp, R. I., Towell, N. A., and Phillips,
and relation to perception. Hum. Neurobiol. 3: 197–208.
K. C. (1997). Recognising moving faces: The relative con-
Perrett, D. I., Smith, P. A. J., Potter, D. D., Mistlin, A. J., tribution of motion and perspective view information.
Head, A. S., Milner, A. D., and Jeeves, M. A. (1985). Visual Vis. Cognit. 4: 409–438.
cells in the temporal cortex sensitive to face view and gaze
Pisoni, D. B. (2000). Cognitive factors and cochlear
direction. Proc. R. Soc. Lond. B Biol. Sci. 223: 293–317.
implants: Some thoughts on perception, learning, and
Peters, A., and Kaiserman-Abramof, I. R. (1969). The memory in speech perception. Ear Hear. 21: 70–78.
small pyramidal neuron of the rat cerebral cortex: The
Pitts, W. and McCulloch, W. S. (1947/1965). How we
synapses upon dendritic spines. Z. Zellforsch. Mikroskop.
know universals: The perception of auditory and visual
Anat. 100: 487–506.
forms. In Embodiments of Mind, 46–66. Reprint, Cam-
Peters, A., Payne, B. R., and Rudd, J. (1994). A numerical bridge, MA: MIT Press, 1965.
analysis of the geniculocortical input to striate cortex in
Pizlo, Z. (1994). A theory of shape constancy based on
the monkey. Cereb. Cortex 4: 215–229.
perspective invariants. Vision Res. 34: 1637–1658.
Peterson, M. A., and Gibson, B. S. (1993). Shape rec-
Plaisted, K., O’Riordan, M., and Baron-Cohen, S. (1998).
ognition inputs to figure-ground organization in three-
Enhanced discrimination of novel, highly similar stimuli by
dimensional display. Cogn. Psychol. 25: 383–429.
adults with autism during a perceptual learning task. J.
Peterson, M. A., and Gibson, B. S. (1994). Must figure- Child Psychol. Psychiatry 39: 765–775.
ground organization precede object recognition? An as-
Pleger, B., Dinse, H. R., Ragert, P., Schwenkreis, P.,
sumption in peril. Psychol. Sci. 5: 253–259.
Malin, J. P., and Tegentho¤, M. (2001). Shifts in cortical
Petry, S., and Meyer, G., eds. (1987). The Perception of Illu- representations predict human discrimination improve-
sory Contours. New York: Springer. ment. Proc. Natl. Acad. Sci. U S A 98: 12255–12260.
Pettet, M. W., and Gilbert, C. D. (1992). Dynamic Poggio, T. (1990). A theory of how the brain might work.
changes in receptive-field size in cat primary visual cortex. Cold Spring Harb. Symp. Quant. Biol. 55: 899–910.
Proc. Natl. Acad. Sci. U S A 89: 8366–8370.
Poggio, T., and Edelman, S. (1990). A network that learns
Pettit, M. J., and Schwark, H. D. (1993). Receptive field to recognize three-dimensional objects. Nature 343: 263–
reorganization in dorsal column nuclei during temporary 266.
denervation. Science 262: 2054–2056.
Poggio, T., Edelman, S., and Fahle, M. (1992). Learning of
Pham, D. T., Garrat, P., and Jutten, C. (1992). Separation visual modules from examples: A framework for under-
of a mixture of independent sources through a maximum standing adaptive visual performance. Comput. Vis. Graph.
likelihood approach. In Proceedings of European Signal Pro- Image Proc. Image Understand. 56: 22–30.
cessing Conference (EUSIPCO), pp. 771–774.
Poggio, T., Fahle, M., and Edelman, S. (1992). Fast percep-
Phillips, P. J. (1998). Foundations of face recognition. In tual learning in visual hyperacuity. Science 256: 1018–1021.
H. Wechsler et al., eds., Face Recognition: From Theory to
Poggio, T., and Girosi, F. (1990). Regularization algo-
Applications. Berlin: Springer.
rithms for learning that are equivalent to multilayer net-
Phillips, R. J. (1972). Why are faces hard to recognise works. Science 247: 978–982.
in photographic negative? Percept. Psychophys. 12: 425–
Poggio, T. A., and Vetter, T. (1992). Recognition and
426.
structure from one 2-D model view: Observations on pro-
426 References
totypes, object classes, and symmetries. AI Memo 1347. Prados, J., Chamizo, V. D., and Mackintosh, N. J. (1999).
MIT AI Laboratory. Latent inhibition and perceptual learning in a swimming-
pool navigation task. J. Exp. Psychol. Anim. Behav. Process.
Polat, U., Mizobe, K., Pettet, M. W., Kasamutsu, T., and
25: 37–44.
Norcia, A. M. (1998). Collinear stimuli regulate visual
responses depending on cell’s contrast threshold. Nature Prakash, N., Cohen-Cory, S., and Frostig, R. D. (1996).
391: 580–584. Rapid and opposite e¤ects of BDNF and NGF on the
functional organization of the adult cortex in vivo. Nature
Polat, U., and Sagi, D. (1993). Lateral interactions between
381: 702–706.
spatial channels: Suppression and facilitation revealed by
lateral masking experiments. Vision Res. 33: 993–999. Prescott, S. W. (1998). Interactions between depression
and facilitation in neural networks: Updating the dual-
Polat, U., and Sagi, D. (1994a). The architecture of per-
process theory of plasticity. Learn. Mem. 5: 446–466.
ceptual spatial interactions. Vision Res. 34: 73–78.
Price, D. J. (1986). The postnatal development of clustered
Polat, U., and Sagi, D. (1994b). Spatial interactions in
intrinsic connections in area 18 of the visual cortex of
human vision: From near to far via experience-dependent
kittens. Dev. Brain Res. 24: 31–38.
cascades of connections. Proc. Natl. Acad. Sci. U S A 91:
1206–1209. Psarrou, A., Gong, S., and Buxton, H. (1995). Modelling
spatio-temporal trajectories and face signatures on partially
Polat, U., and Sagi, D. (1995). Plasticity of spatial in-
recurrent networks. (1995). In Proceedings of the Interna-
teractions in early vision. In B. Julesz and I. Kovács,
tional Conference on Neural Networks: ICNN ’95, pp. 2226–
eds., Maturational Windows and Adult Cortical Plasticity. SFI
2231.
Studies in the Sciences of Complexity, vol. 24. Reading
MA: Addison-Wesley. Purves, D., Augustine, G. J., Fitzpatrick, D., Katz, L. C.,
LaMantia, A.-S., and McNamara, J. O., eds. (1997). Neu-
Pollen, D. A., and Ronner, S. F. (1981). Phase relation-
roscience. Sunderland, MA: Sinauer.
ships between adjacent simple cells in the visual cortex.
Science 212: 1409–1410. Pustell, T. E. (1957). The experimental induction of per-
ceptual vigilance and defense. J. Pers. 25: 425–438.
Polley, D. B., Chen-Bee, C. H., and Frostig, R. D.
(1999a). Varying the degree of single-whisker stimulation Pylyshyn, Z. W. (1984). Computation and Cognition: Toward
di¤erentially a¤ects phases of intrinsic signals in rat barrel a Foundation for Cognitive Science. Cambridge, MA: MIT
cortex. J. Neurophysiol. 81: 692–701. Press.
Polley, D. B., Chen-Bee, C. H., and Frostig, R. D. (1999b). Pylyshyn, Z. W. (1999). Is vision continuous with cogni-
Two directions of plasticity in the sensory-deprived adult tion? The case for cognitive impenetrability of visual per-
cortex. Neuron 24: 623–637. ception Behav. Brain Sci. 22(3): 341–423.
Pollock, W., and Chapais, A. (1952). The apparent length Qian, N., and Matthews, N. (1999). A physiological theory
of a line as a function of its inclination. Q. J. Exp. Psychol. for visual perceptual learning of orientation discrimination.
4: 170–178. Soc. Neurosci. Abstr. 25: 1316.
Pons, T., Garraghty, P. E., Ommaya, A. K., Kaas, J. H., Quine, W. V. O. (1969). Natural kinds. In Ontological
Taub, E., and Mishkin, M. (1991). Massive cortical reor- Relativity and Other Essays, 114–138. New York: Columbia
ganization after sensory dea¤erentiation in adult macaques. University Press.
Science 252: 1857–1860.
Rainer, G., and Miller, E. K. (2000) E¤ects of visual ex-
Popper, K. R. (1992). Conjectures and Refutations: The perience on the representation of objects in the prefrontal
Growth of Scientific Knowledge. 5th ed. London: Routledge. cortex. Neuron 27: 179–189.
Posner, M. I., and Keele, S. W. (1968). On the genesis of Rajan, R., Irvine, D. R. F., Wise, L. Z., and Heil, R.
abstract ideas. J. Exp. Psychol. 77: 353–363. (1993). E¤ect of unilateral partial cochlear lesions in adult
References 427
cats on the representation of lesioned and unlesioned co- Recanzone, G. H., Merzenich, M. M., and Schreiner, C.
chleas in primary auditory centers. J. Comp. Neurol. 338: E. (1992e). Changes in the distributed temporal response
17–49. properties of SI cortical neurons reflect improvements in
performance on a temporally based tactile discrimination
Ramachandran, V. (1985). The neurobiology of percep-
task. J. Neurophysiol. 67: 1071–1091.
tion. Perception 14: 97–103.
Recanzone, G. H., Schreiner, C. E., and Merzenich, M. M.
Ramachandran, V. (1988). Perception of shape from
(1993). Plasticity in the frequency represenation of pri-
shading. Nature 331: 163–166.
mary auditory cortex following discrimination training in
Ramachandran, V. (1994). 2D or not 2D—that is the adult owl monkeys. J. Neurosci. 13: 87–103.
question. In R. Gregory and J. Harris, eds., The Artful Eye.
Redlich, A. N., Atick, J. J., and Gri‰n, P. A. (1996). Sta-
Oxford: Oxford University Press.
tistical approach to shape from shading: Deriving 3-D face
Ramachandran, V. S. (1976). Learning-like phenomena in surfaces from single 2-D images. Network Comput. Neural
stereopsis. Nature 262: 382–384. Syst. 7: 1.
Ramachandran, V. S., and Braddick, O. (1973). Orienta- Regan, D., and Beverley, K. I. (1985). Postadaptation
tion-specific learning in stereopsis. Perception 2: 371–376. orientation discrimination. J. Opt. Soc. Am. A2: 147–
Ramachandran, V. S., Stewart, M., and Rogers- 155.
Ramachandran, D. C. (1992). Perceptual correlates of Reicher, G. M. (1969). Perceptual recognition as a func-
massive cortical reorganization. NeuroReport 3: 583–586. tion of meaningfulness of stimulus material. J. Exp. Psychol.
Rasmusson, D. D., and Dykes, R. W. (1988). Long-term 81: 275–280.
enhancement of evoked potentials in cat somatosensory Reinke, H., and Dinse, H. R. (1999). Plasticity in the
cortex produced by co-activation of the basal forebrain somatosensory and motor cortex of rats: impact of age and
and cutaneous receptors. Exp. Brain Res. 70: 276–286. housing conditions. In Proceedings of the First Göttingen
Recanzone, G. H., Jenkins, W. M., Hradek, G. T., and Conference of the German Neuroscience Society 1999: From
Merzenich, M. M. (1992a). Progressive improvements in Molecular Neurobiology to Clinical Neuroscience, ed. N. Elsner
discriminative abilities in adult owl monkeys performing a and U. Eysel, vol. 1, pp. 409. Stuttgart: Thieme.
tactile frequency discrimination task. J. Neurophysiol. 67: Rentschler, I., Jüttner, M., and Caelli, T. (1994). Proba-
1015–1030. bilistic analysis of human supervised learning and classifi-
Recanzone, G. H., Merzenich, M. M., and Dinse, H. R. cation. Vision Res. 34: 669–687.
(1992b). Expansion of the cortical representation of a spe- Rhodes, G. (1996). Superportraits: Caricature and Recogni-
cific skin field in primary somatosensory cortex by intra- tion. Hove, U.K.: Psychology Press.
cortical microstimulation. Cereb. Cortex 2: 181–196.
Rhodes, G., Brake, S., and Atkinson, A. (1993). What’s
Recanzone, G. H., Merzenich, M. M., and Jenkins, W. M. lost in inverted faces? Cognition 47: 25–57.
(1992c). Frequency discrimination training engaging a re-
Rhodes, G., Brennan, S., and Carey, S. (1987). Identifi-
stricted skin surface results in an emergence of a cutaneous
cation and ratings of caricatures: Implications for mental
response zone in cortical area 3a. J. Neurophysiol. 67:
representations of faces. Cogn. Psychol. 19: 473–497.
1057–1070.
Rhodes, G., Carey, S., Byatt, G., and Pro‰tt, F. (1998).
Recanzone, G. H., Merzenich, M. M., Jenkins, W. M.,
Coding spatial variations in faces and simple shapes: A test
Grajski, K., and Dinse, H. R. (1992d). Topographic reor-
of two models. Vision Res. 38: 15–16.
ganization of the hand representation in cortical area 3b of
owl monkeys trained in a frequency discrimination task. J. Riches, I. P., Wilson, F. A. W., and Brown, M. W. (1991).
Neurophysiol. 67: 1031–1056. The e¤ects of visual stimulation and memory on neurons
428 References
of the hippocampal formation and the neighboring para- Rockland, K. S., and Lund, J. S. (1982). Widespread peri-
hippocampal gyrus and inferior temporal cortex of the odic intrinsic connections in the tree shrew visual cortex.
primate. J. Neurosci. 11: 1763–1779. Science 215: 1532–1534.
Richmond, B. J., and Optican, L. M. (1987). Temporal Rockland, K. S., and Lund, J. S. (1983). Intrinsic laminar
encoding of two-dimensional patterns by single units in lattice connections in primate visual cortex. J. Comp.
primate inferior temporal cortex: 2. Quantification of re- Neurol. 216: 303–318.
sponse waveform. J. Neurophysiol. 57: 147–161.
Röder, B., Rösler, F., and Neville, H. J. (1999). E¤ects of
Richmond, B. J., Optican, L. M., Podell, M., and Spitzer, interstimulus interval on auditory event-related potentials
H. (1987). Temporal encoding of two-dimensional in congenitally blind and normally sighted humans. Neu-
patterns by single units in primate inferior temporal cor- rosci. Lett. 264: 53–56.
tex. 1. Response characteristics. J. Neurophysiol. 57: 132–
Röder, B., Rösler, F., and Neville, H. J. (2000). Event-
146.
related potentials during language processing in con-
Ridding, M. C., and Rothwell, J. C. (1995). Reorganiza- genitally blind and sighted people. Neuropsychologia 38:
tion in human motor cortex. Can. J. Physiol. Pharmacol. 73: 1482–1502.
218–222.
Röder, B., Rösler, F., Henninghausen, E., and Nacker, F.
Riesenhuber, M., and Poggio, T. (1998). Just one view: (1996). Event-related potentials during auditory and soma-
Invariances in inferotemporal cell tuning. In M. I. Jordan, tosensory discrimination in sighted and blind human sub-
M. J. Kearns, and S. A. Solla, eds., Advances in Neural In- jects. Cogn. Brain Res. 4(2): 77–93.
formation Processing, vol. 10, pp. 215–221. Cambridge, MA:
Röder, B., Teder-Sälejärvi, W., Sterr, A., Rösler, F.,
MIT Press.
Hillyard, S. A., and Neville, H. (1999). Improved audi-
Ringach, D., and Shapley, R. (1996). Spatial and temporal tory spatial tuning in blind humans. Nature 400: 162–166.
properties of illusory contours and amodal boundary com-
Rösler, F., Röder, B., Heil, M., and Henninghausen, E.
pletion. Vision Res. 36: 3037–3050.
(1993). Topographic di¤erences of slow event-related brain
Rivenes, R. S. (1967). Multiple-task transfer e¤ects in potentials in blind and sighted adult human subjects during
perceptual-motor learning. Res. Q. 38: 485–493. haptic mental rotation. Cogn. Brain Res. 1: 145–159.
Rivest, J., Boutet, I., and Intriligator, J. (1997). Perceptual Rodman, H. R., and Consuelos, M. J. (1994). Cortical
learning of orientation discrimination by more than one projects to anterior inferior temporal cortex in infant
attribute. Vision Res. 37: 273–281. macaque monkeys. Vis. Neurosci. 11: 119–133.
Roberson, E. D., English, J. D., and Sweatt, J. D. (1996). Rodman, H. R., O Scalaidhe, S. P., and Gross, C. G.
A biochemist’s view of long-term potentiation. Learn. (1993). Response properties of neurons in temporal corti-
Mem. 3: 1–24. cal visual areas of infant monkeys. J. Neurophysiol. 70: 1115–
1136.
Robertson, D., and Irvine, D. R. F. (1989). Plasticity of
frequency organization in auditory cortex of guinea pigs Roelfsema, P. R., Lamme, V. A. F., and Spekreijse, H.
with partial unilateral deafness. J. Comp. Neurol. 282: 456– (1998). Object-based attention in the primary visual cortex
471. of the macaque monkey. Nature 395: 376–381.
Rock, I., and DiVita, J. (1987). A case of viewer-centered Rolls, E. (1992). Neurophysiological mechanisms under-
object perception. Cogn. Psychol. 19: 280–293. lying face processing within and beyond the temporal cortical
areas. Philos. Trans. R. Soc. Lond. B Biol. Sci. 335: 11–21.
Rockland, K. S. (1985). Anatomical organization of pri-
mary visual cortex (area 17) in the ferret. J. Comp. Neurol. Rolls, E. T. (1984). Neurons in the cortex of the temporal
241: 225–236. lobe and in the amygdala of the monkey with responses
selective for faces. Hum. Neurobiol. 3: 209–222.
References 429
Rolls, E. T. (1994). Brain mechanisms for invariant visual Rossini, P. M., Martino, G., Narici, L., Pasquarelli, A.,
recognition and learning. Behav. Processes 33: 113–138. Peresson, M., Pizzella, V., Tecchio, F., Torrioli, G., and
Romani, G. L. (1994). Short-term brain ‘‘plasticity’’ in
Rolls, E. T., and Baylis, G. C. (1986). Size and contrast
humans: Transient finger representation changes sensory
have only small e¤ects on the responses to faces of neurons
cortex somatotopy following ischemic anesthesia. Brain
in the cortex of the superior temporal sulcus of the monkey.
Res. 642: 169–177.
Exp. Brain Res. 65: 38–48.
Rovamo, J., and Virsu, V. (1979). An estimation and ap-
Rolls, E. T., Baylis, G. C., Hasselmo, M. E., and Nalwa,
plication of the human cortical magnification factor. Exp.
V. (1989). The e¤ect of learning on the face selective
Brain Res. 37: 1–20.
responses of neurons in the cortex in the superior tempo-
ral sulcus of the monkey. Exp. Brain Res. 76: 153–164. Rubenstein, B. S., and Sagi, D. (1990). Spatial variability
as a limiting factor in texture discrimination tasks: Impli-
Rolls, E. T., Baylis, G. C., and Leonard, C. M. (1985). Role
cation for performance asymmetries. J. Opt. Soc. Am. A 7:
of low and high spatial frequencies in the face-selective
1632–1643.
responses of neurons in the cortex in the superior tempo-
ral sulcus. Vision Res. 25: 1021–1035. Rubin, N., Nakayama, K., and Shapley, R. (1996). En-
hanced perception of illusory contours in the lower versus
Rolls, E. T., and Tovee, M. J. (1994). Processing speed
upper visual hemifields. Science 271: 651–653.
in the cerebral cortex and the neurophysiology of visual
masking. Proc. R. Soc. Lond. B Biol. Sci. 257: 9–15. Rubin, N., Nakayama, K., and Shapley, R. (1997). Abrupt
learning and retinal size specifity in illusory-contour per-
Rolls, E. T., and Tovee, M. J. (1995). Sparseness of the
ception. Curr. Biol. 7: 461–467.
neuronal representation of stimuli in the primate temporal
visual cortex. J. Neurophysiol. 73: 713–726. Ruderman, D. L. (1994). The statistics of natural images.
Network 5: 517–548.
Rosch, E. (1973). On the internal structure of perceptual
and semantic categories. In T. Moore, ed., Cognitive De- Rumelhart, D. E., and McClelland, J. L., eds. (1986). Par-
velopment and the Aquisition of Language, 111–144. New allel Distributed Processing. Cambridge, MA: MIT Press.
York: Academic Press.
Rumelhart, D. E., and Todd, P. M. (1993). Learning and
Rosch, E., Mervis, C. B., Gray, W. D., Johnson, D. M., connectionist representations. In D. E. Meyer and S.
and Boyes-Braem, P. (1976). Basic objects in natural cate- Kornblum, eds., Attention and Performance XIV, 3–34.
gories. Cogn. Psychol. 8: 382–439. Cambridge, MA: MIT Press.
Rosenthal, O., and Hochstein, S. (1994). E¤ects of stim- Rumpel, S., Ho¤mann, H., Gottmann, K., Hatt, H.,
ulus meaning on V1 orientation-tuning properties. Isr. Soc. Mittmann, T., and Eysel, U. T. (1999). Lesion-induced
Neurosci. 3: 47. changes in transcription levels of NMDA receptor subunit
mRNAs in visual cortex of rats. Soc. Neurosci. Abstr. 25:
Rosier, A. M., Arckens, L., Demeulemeester, H., Orban,
1719.
G. A., Eysel, U. T., Wu, Y.-J., and Vandesande, F. (1995).
E¤ect of sensory dea¤erentation on immunoreactivity of Russell, G. (1976). Practice e¤ects for auditory localiza-
GABAergic cells and on GABA receptors in the adult cat tion: A test of a di¤erentiation theory of perceptual learn-
visual cortex. J. Comp. Neurol. 359: 476–489. ing and development. Percept. Mot. Skills 42: 647–653.
Ross, H. (1990). Environmental influences on geometrical Ruthazer, E. S., and Stryker, M. P. (1996). The role of
illusions. In F. Muller, ed., Fechner Day ’90: Proceedings of activity in the development of long-range horizontal con-
the Sixth Annual Meeting of the International Society of Psy- nections in area 17 of the ferret. J. Neurosci. 16: 7253–
chophysicists, 216–221. 7269.
Ross, H., and Woodhouse, J. (1979). Genetic and envi- Saarinen, J., and Levi, D. M. (1995). Perceptual learning in
ronmental factors in orientation anisotropy: A field study vernier acuity: What is learned? Vision Res. 35: 519–527.
in the British Isles. Perception 8: 507–521.
430 References
Sachdev, R. N., Lu, S. M., Wiley, R. G., and Ebner, F. F. Saksida, L. M. (1999). E¤ects of similarity and experience
(1998). Role of the basal forebrain cholinergic projection on discrimination learning: A nonassociative connectionist
in somatosensory cortical plasticity. J. Neurophysiol. 79: model of perceptual learning. J. Exp. Psychol. Anim Behav.
3216–3228. Process. 25: 308–323.
Sadato, N., Pascual-Leone, A., Grafman, J., Deiber, M.-P., Saleem, K. S., and Tanaka, K. (1996). Divergent projec-
Ibanez, V., and Hallett, M. (1998). Neural networks for tions from the anterior inferotemporal area TE to the peri-
Braille reading in the blind. Brain 121: 1213–1229. rhinal and entorhinal cortices in the macaque monkey. J.
Neurosci. 16: 4757–4775.
Sadato, N., Pascual-Leone, A., Grafman, J., Ibanez, V.,
Deiber, M. P., Dold, G., and Hallett, M. (1996). Activa- Saleem, K. S., Tanaka, K., and Rockland, K. S. (1993).
tion of the primary visual cortex by Braille reading in blind Specific and columnar projection from area TEO to TE in
subjects. Nature 380: 526–528. the macaque inferotemporal cortex. Cereb. Cortex 3: 454–
464.
Sa¤ran, J. R., Aslin, R. N., and Newport, E. L. (1996).
Statistical learning by 8-month-old infants. Science 274: Sameshima, K., and Merzenich, M. M. (1993). Cortical
1926–1928. plasticity and memory. Curr. Opin. Neurobiol. 3: 187–196.
Sagi, D. (1990). Detection of an orientation singularity Sanes, J. N., and Donoghue, J. P. (1997). Static and dy-
in Gabor textures: E¤ect of signal density and spatial fre- namic organization of motor cortex. Adv. Neurol. 73: 277–
quency. Vision Res. 30: 1377–1388. 296.
Sagi, D. (1996). Early vision: Images, context and mem- Sanger, T. D. (1989). Optimal unsupervised learning in a
ory. In A. Aertsen and V. Braitenberg, eds., Brain Theory: single-layer network. Neural Networks 2: 459–473.
Biological Basis and Computational Theory of Vision, 1–15.
Sarle, W. S. (1994). Neural networks and statistical
Amsterdam: Elsevier.
models. In Proceedings of the Nineteenth Annual SAS Users
Sagi, D., and Hochstein, S. (1985). Lateral inhibition be- Group International Conference, 1538–1550, Cary, NC: SAS
tween spatially adjacent spatial frequency channels? Percept. Institute.
Psychophys. 37: 315–322.
Sary, G., Vogels, R., and Orban, G. A. (1993). Cue-
Sagi, D., and Julesz, B. (1987). Short-range limitation on invariant shape selectivity of macaque inferior temporal
detection of feature di¤erences. Spat. Vis. 2: 39–49. neurons. Science 260: 995–997.
Sagi, D., and Tanne, D. (1994). Perceptual learning: Saslow, M. G. (1967). Latency for saccadic eye movement.
learning to see. Curr. Opin. Neurobiol. 4: 195–199. J. Opt. Soc. Am. 57: 1030–1036.
Saito, S., Kobayashi, S., Ohashi, Y., Igarashi, M., Komiya, Sathian, K., and Zangaladze, A. (1997). Tactile learning is
Y., and Ando, S. (1994). Decreased synaptic density in task specific but transfers between fingers. Percept. Psycho-
aged brains and its prevention by rearing under enriched phys. 59: 119–128.
environment as revealed by synaptophysin contents. J.
Sathian, K., and Zangaladze, A. (1998). Perceptual learn-
Neurosci. Res. 39: 57–62.
ing in tactile hyperacuity: Complete intermanual transfer
Sakai, K., and Miyashita, Y. (1991). Neural organization but limited retention. Exp. Brain Res. 118: 131–134.
for the long-term memory of paired associates. Nature 354:
Sato, T. (1988). E¤ects of attention and stimulus interac-
152–155.
tion on visual responses of inferior temporal neurons in
Sakai, K., and Miyashita, Y. (1994). Neuronal tuning to macaque. J. Neurophysiol. 60: 344–364.
learned complex forms in vision. NeuroReport 5: 829–832.
Sato, T. (1989). Interactions of visual stimuli in the recep-
Sakai, K., Naya, Y., and Miyashita, Y. (1994). Neuronal tive fields of inferior temporal neurons in awake macaques.
tuning and associative mechanisms in form representation. Exp. Brain Res. 77: 23–30.
Learn. Mem. 1: 83–105.
References 431
Sato, T. (1995). Interactions between two di¤erent visual Schmidt, K. E., Kim, D.-S., Singer, W., Bonhoe¤er, T.,
stimuli in the receptive fields of inferior temporal neurons and Löwel, S. (1997). Functional specificities of long-range
in macaques during matching behaviors. Exp. Brain Res. and interhemispheric connections in the visual cortex of
105: 209–219. strabismic cats. J. Neurosci. 17: 5480–5492.
Scannell, J. W., and Young, M. P. (1999). Neuronal pop- Schmidt-Kastner, R., Wietasch, K., Weigel, H., and
ulation activity and functional imaging. Proc. R. Soc. Lond. Eysel, U. T. (1993). Immunohistochemical staining for
B Biol. Sci. 266: 875–881. glial fibrillary acidic protein (GFAP) after dea¤erentation or
ischemic infarction in rat visual system: Features of reactive
Schacter, D. L., Cooper, L. A., and Delaney, S. M. (1990).
and damaged astrocytes. Int. J. Dev. Neurosci. 11: 157–174.
Implicit memory for unfamiliar objects depends on access
to structural descriptions. J. Exp. Psychol. Gen. 119: 5–24. Schooler, J. W., Fallshore, M., and Fiore, S. M. (1995).
Epilogue: Putting insight into perspective. In The Nature of
Schafer, R., and Murphy, G. (1943). The role of autism
Insight, ed. R. J. Sternberg and J. E. Davidson. Cambridge,
in a visual figure-ground relationship. J. Exp. Psychol. 32:
MA: MIT Press.
335–343.
Schoups, A. A., and Orban, G. A. (1996). Interocular
Sche¿er, K., Bilecen, D., Schmid, N., Tschopp, K., and
transfer in perceptual learning of a pop-out discrimination
Seelig, J. (1998). Auditory cortical responses in hearing sub-
task. Proc. Natl. Acad. Sci. U S A 93: 7358–7362.
jects and unilateral deaf patients as detected by functional
magnetic resonance imaging. Cereb. Cortex 8: 156–163. Schoups, A. A., Vogels, R., and Orban, G. A. (1995).
Human perceptual learning in identifying the oblique ori-
Schiene, K., Bruehl, C., Zilles, K., Qü, M., Hagemann, G.
entation: Retinotopy, orientation specificity and mono-
Kraemer, M., and Witte, O. W. (1996). Neuronal hyper-
cularity. J. Physiol. (Lond.) 483: 797–810.
excitability and reduction of GABAA -receptor expression
in the surround of cerebral photothrombosis. J. Cereb. Schoups, A. A., Vogels, R., and Orban, G. A. (1998).
Blood Flow Metab. 16: 906–914. E¤ects of perceptual learning in orientation discrimination
on orientation coding in V1. Invest. Ophthalmol. Vis. Sci.
Schiltz, C., Bodart, J. M., Dubois, S., Dejardin, S., Michel,
39: S684: 3142.
C., Roucoux, A., Crommelinck, and Orban, G. A. (1999).
Neuronal mechanisms of perceptual learning: Changes in Schoups, A. A., Vogels, R., Qian, N., and Orban, G. A.
human brain activity with training in orientation discrimi- (2001). Practising orientation identification improves ori-
nation. Neuroimage 9: 46–62. entation coding in V1 neurons. Nature 412: 549–553.
Schlaug, G., Jancke, L., Huang, Y., and Steinmetz, H. Schroeter, M., Schiene, K., Kraemer, M., Hagemann, G.,
(1995). In-vivo evidence of structural brain asymmetry in Weigel, H., Eysel, U. T., Witte, O. W., and Stoll, G.
musicians. Science 267: 699–701. (1995). Astroglial responses in photochemically induced
focal ischemia of the rat cortex. Exp. Brain Res. 106: 1–6.
Schmid, L. M., Rosa, M. G. P., Calford, M. B., and
Ambler, J. S. (1996). Visuotopic reorganization in the pri- Schuurman, T., Klein, H., Beneke, M., and Traber, J.
mary cortex of adult cats following monocular and binoc- (1987). Nimodipine and motor deficits in the aged rat.
ular retinal lesions. Cereb. Cortex 6: 388–405. Neurosci. Res. Comm. 1: 9–15.
Schmidhuber, J., Eldracher, M., and Foltin, B. (1996). Schwartz, E. L., Desimone, R., Albright, T. D., and Gross,
Semi-linear predictability minimization produces well- C. G. (1983). Shape recognition and inferior temporal
known feature detectors. Neural Comput. 8: 773–86. neurons. Proc. Natl. Acad. Sci. U S A 80: 5776–5778.
Schmidt, K. E., Goebel, R., Löwel, S., and Singer, W. Schwartz, O., and Simoncelli, E. P. (1999). Accounting
(1997). The perceptual grouping criterion of colinearity for surround suppression in V1 neurons using a statistically
is reflected by anisotropies of connections in the primary optimized normalization model. Invest. Ophthalmol. Vis.
visual cortex. Eur. J. Neurosci. 9: 1083–1089. Sci. 40: S641.
432 References
Schwarz, C., and Bolz, J. (1991). Functional specificity of a Shannon, C. E. (1948). A mathematical theory of com-
long-range horizontal connection in cat visual cortex: A munication. Bell Sys. Tech. J. 27: 623–656.
cross-correlation study. J. Neurosci. 11: 2995–3007.
Shapiro, P. N., and Penrod, S. (1986). Meta-analysis
Schweigart, G., and Eysel, U. T. (1998). Receptive fields of facial identification studies. Psychol. Bull. 100: 139–
near to excitotoxic lesions in the visual cortex of the cat. 156.
Soc. Neurosci. Abstr. 24: 647.
Shapley, R., and Enroth-Cugell, C. (1984). Visual adapta-
Schweigart, G., and Eysel, U. T. (Forthcoming). Activity- tion and retinal gain controls. In Progress in Retinal Research,
dependent receptive field changes in the surround of adult ed. J. G. Osborn and N. N. Chadler, vol. 3, pp. 263–346.
cat visual cortex lesions. Oxford: Pergamon Press.
Schweinberger, S. R., and Soukoup, G. R. (1998). Shashua, A., and Ullman, S. (1988). Structural saliency.
Asymmetric relationships between the perception of facial The detection of globally salient structures using a locally
identity, emotion and facial speech. J. Exp. Psychol. Hum. connected network. In Proceedings of the International Con-
Percept. Perform. 24: 1748–1765. ference on Computer Vision. Los Alamitos, CA: IEEE Press.
Schyns, P. (1998). Categories and percepts: A bi-directional Shatz, C. J., Lindström, S., and Wiesel, T. N. (1977). The
framework for categorization. Trends Cogn. Sci. 1: 183– distribution of a¤erents representing the right and left eyes
189. in the cat’s visual cortex. Brain Res. 131: 103–116.
Schyns, P. G., and Rodet, L. (1997). Categorization cre- Sheinberg, D. L., and Logothetis, N. K. (1997). The role
ates function features. J. Exp. Psychol. Learn. Mem. Cogn. of temporal cortical areas in perceptual organization. Proc.
23: 681–696. Natl. Acad. Sci. U S A 94: 3408–3413.
Schyns, P., Goldstone, R., and Thibaut, J.-P. (1998). Shenoy, K. V., Kaufman, J., McGrann, J. V., and Shaw,
The development of features in object concepts. Behav. G. L. (1993). Learning by selection in the trion model
Brain Sci. 21: 1–54. of cortical organization. Cereb. Cortex 3: 239–248.
Scott, D. V. (1974). Perceptual learning. Queens Nurs J. Shepard, R. N. (1967). Recognition memory for words,
17: 2–3. sentences and pictures. J. Verb. Learn. Verb. Behav. 6: 156–
163.
Searcy, J. H., and Bartlett, J. C. (1996). Inversion and
processing of component and spatial-relational information Shepard, R. N. (1987). Toward a universal law of gener-
in faces. J. Exp. Psychol. Hum. Percept. Perform. 22: 904– alization for psychological science. Science 237: 1317–
915. 1323.
Seitz, R. J., Huang, Y., Knorr, U., Tellmann, L., Herzog, Shepard, R., and Cooper, L. (1982). Mental Images and
H., and Freund, H.-J. (1995). Large-scale plasticity in the Their Transforms. 3d ed. Cambridge, MA: MIT Press.
human motor cortex. NeuroReport 6(5): 742–744.
Shepherd, G. M. (1990). The significance of real neuron
Selfridge, O. G. (1959). Pandemonium: A paradigm for architectures for neural network simulations. In E. L.
learning. In The Mechanisation of Thought Processes. London: Schwartz, ed., Computational Neuroscience. Cambridge, MA:
H.M.S.O. MIT Press.
Sengpiel, F., Baddeley, R. J., Freeman, T. C., Harrad, R., Shepherd, J., Davies, G., and Ellis, H. (1981). Studies of
and Blakemore, C. (1998). Di¤erent mechanisms underlie cue saliency. In G. Davies, H. Ellis, and J. Shepherd, eds.,
three inhibitory phenomena in cat area 17. Vision Res. Perceiving and Remembering Faces, pp. 105–131. London:
38(14): 2067–2080. Academic Press.
Sengpiel, F., Stawinski, P., and Bonhoe¤er, T. (1999). Shiu, L., and Pashler, H. (1992). Improvement in line
Influence of experience on orientation maps in cat visual orientation discrimination is retinally local but dependent
cortex. Nat. Neurosci. 2: 727–732. on cognitive set. Percept. Psychophys. 52: 582–588.
References 433
Shouval, H. (1995). Formation and organisation of recep- Skrandies, W., and Jedynak, A. (1999). Learning to see 3–
tive fields, with an input environment composed of natural D: Psychophysics and brain electrical activity. NeuroReport
scenes. Ph.D. dissertation, Brown University. 10: 249–253.
Sil’kis, I. G., and Rapoport, S. S. (1995). Plastic reorgani- Skrandies, W., Jedynak, A., and Fahle, M. (2001). Per-
zations of the receptive fields of neurons of the auditory ceptual learning: Psychophysical thresholds and electrical
cortex and the medial geniculate body induced by micro- brain topography. Int. J. Psychophysiol. 41: 119–129.
stimulation of the auditory cortex. Neurosci. Behav. Physiol.
Skrandies, W., Lang, G., and Jedynak, A. (1996). Sensory
25: 322–339.
thresholds and neurophysiological correlates of human
Singer, W. (1990). Search for coherence: A basic principle perceptual learning. Spat. Vis. 9: 475–489.
of cortical self-organization. Concepts Neurosci. 1: 1–26.
Smith, D. E. P., and Hochberg, J. E. (1954). The e¤ect of
Singer, W. (1995). Development and plasticity of cortical ‘‘punishment’’ (electric shock) on figure-ground percep-
processing architectures. Science 270: 758–764. tion. J. Psychol. 38: 83–87.
Singer, W., Engel, A. K., Kreiter, A. K., Munk, M. H. J., Smith, V. C., and Pokorny, J. (1975). Spectral sensitivity
Neuenschwander, S., and Roelfsema, P. R. (1997). Neu- of the foveal cone photopigments between 400 and 500
ronal assemblies: necessity, signature and detectability. nm. Vision Res. 15: 161–171.
Trends Cogn. Sci. 1: 252–261.
Sober, S. J., Stark, J. M., Yamasaki, D. S., and Lytton,
Singer, W., and Tretter, F. (1976). Unusually large recep- W. W. (1997). Receptive field changes after strokelike cor-
tive fields in cats with restricted visual experience. Exp. tical ablation: A role for activation dynamics. J. Neuro-
Brain Res. 26: 171–184. physiol. 78: 3438–3443.
Sinha, P. (1994). Object recognition via image invariants. Sokolo¤, L., Reivich, M., Kennedy, C., DesRosiers, M.
Invest. Ophthalmol. Vis. Sci. 35(4): 1735. H., Patlak, C. S., Pettigrew, K. D., Sakurada, O., and
Shinohara, M. (1977). The [ 14 C]deoxyglucose method for
Sinha, P. (1995). Perceiving and recognizing three-
the measurement of local cerebral glucose utilization:
dimensional forms. Ph.D. diss., Massachusetts Institute of
Theory, procedure, and normal values in the conscious
Technology.
and anesthetized albino rat. J. Neurochem. 28: 897–916.
Sinha, P., and Adelson, E. H. (1993). Recovering reflec-
Solso, R., and McCarthy, J. (1981). Prototype formation
tance and illumination in a world of painted polyhedra. In
of faces: A case of pseudo-memory. Br. J. Psychol. 72: 499–
Proceedings of the IEEE International Conference on Computer
503.
Vision. Los Alamitos, CA: IEEE Press.
Somers, D. C., Todorov, E. V., Siapas, A. G., Toth, L. J.,
Sinha, P., and Poggio, T. A. (1996). Role of learning in
Kim, D. S., and Sur, M. (1998). A local circuit approach to
three-dimensional form perception. Nature 384: 460–463.
understanding integration of long-range inputs in primary
Sireteanu, R., and Rettenbach, R. (1995). Perceptual visual cortex. Cerebr. Cortex 8: 204–217.
learning in visual search: Fast, enduring, but non-specific.
Sowden, P. T. (1995). On perceptual learning. PhD thesis,
Vision Res. 35: 2037–2043.
University of Surrey, UK
Sireteanu, R., and Rettenbach, R. (2000). Perceptual
Sowden, P. T., Davies, I. R., and Roling, P. (2000). Per-
learning in visual search generalizes over tasks, locations,
ceptual learning of the detection of features in X-ray
and eyes. Vision Res. 40: 2925–2949.
images: A functional role for improvements in adults’
Skrandies, W. (1995). Visual information processing: topo- visual sensitivity? J. Exp. Psychol. Hum. Percept. Perform. 26:
graphy of brain electrical activity. Biol. Psychol. 40: 1–15. 379–390.
Skrandies, W., and Fahle, M. (1994). Neurophysiological Sowden, P., Davies, I., Rose, D., and Kaye, M. (1996).
correlates of perceptual learning in the human brain. Brain Perceptual learning of stereoacuity. Perception 25: 1043–
Topog. 7: 163–168. 1052.
434 References
Spelke, E. (1990). Origins of visual knowledge. In D. Stevenage, S. V. (1998). Which twin are you? A demon-
Osherson and S. M. Kosslyn, eds., Visual Cognition and stration of induced categorical perception of identical twin
Action: An Invitation to Cognitive Science, vol. 2, pp. 99–127. faces. Br. J. Psychol. 89: 39–58.
Cambridge, MA: MIT Press.
Stickgold, R., Whidbee, D., Schirmer, B., Patel, V., and
Spengler, F., and Dinse, H. R. (1994). Reversible reloca- Hobson, J. A. (2000). Visual discrimination task improve-
tion of representational boundaries of adult rats by intra- ment: A multi-step process occurring during sleep. J. Cogn.
cortical microstimulation (ICMS). NeuroReport 5: 949–953. Neurosci. 12: 246–254.
Spengler, F., Godde, B., and Dinse, H. R. (1995). E¤ects Stoll, S., Dorner, H., Blösch, M., and Platt, R. (1990).
of aging on topographic organization of somatosensory Age-dependent di¤erences in the gait of rats. Arch. Gerontol.
cortex. NeuroReport 6: 469–473. Geriatr. 10: 216–268.
Spengler, F., Roberts, T. P., Poeppel, D., Byl, N., Wang, Stone, C. J. (1980). Optimal rates of convergence for
X., Rowley, H. A., and Merzenich, M. M. (1997). nonparametric estimators. Ann. Stat. 8: 1348–1360.
Learning transfer and neuronal plasticity in humans trained
Stone, C. J. (1982). Optimal global rates of convergence
in tactile discrimination. Neurosci. Lett. 232: 151–154.
for nonparametric regression. Ann. Stat. 10: 1040–1053.
Spitzer, H., Desimone, R., and Moran, J. (1988). Increased
Stone, C. J. (1985). Additive regression and other non-
attention enhances both behavioral and neuronal perfor-
parametric models. Ann. Stat. 13: 689–705.
mance. Science 240: 338–340.
Stone, C. J. (1986). The dimensionality reduction principle
Squire, L. R., and Zola, S. M. (1997). Amnesia, memory
for generalized additive models. Ann. Stat. 14: 590–606.
and brain systems. Philos. Trans. R. Soc. Lond. B Biol. Sci.
352: 1663–1673. Stone, J. (1998). Object recognition using spatio-temporal
signatures. Vision Res. 38(7): 947–951.
Stadler, M. A. (1989). On learning complex procedural
knowledge. J. Exp. Psychol. Learn. Mem. Cogn. 15: 1061– Stone, J. V. (1996). Learning perceptually salient visual
1069. parameters using spatiotemporal smoothness constraints.
Neural Comput. 8: 1463–1492.
Standing, L. (1973). Learning 10,000 pictures. Q. J. Exp.
Psychol. 25: 207–222. Stone, J. V. (1996). A canonical microfunction for learning
perceptual invariances. Perception 25: 207–220.
Stemmler, M., Usher, M., and Niebur, E. (1995). Lateral
interactions in primary visual cortex: A model bridging Stone, J. V., and Harper, N. (1999). Temporal constraints
physiology and psychophysics. Science 269: 1877–1880. on visual learning: A computational model. Perception 28:
1089–1104.
Stent, G. S. (1973). A physiological mechanism for Hebb’s
postulate of learning. Proc. Natl. Acad. Sci. U S A 70: 997– Stoney, S. D., Jr., Thompson, W. D., and Asanuma, H.
1001. (1968). Excitation of pyramidal tract cells by intracortical
microstimulation: E¤ective extent of stimulating current.
Sternberg, R. J., and Davidson, J. E., eds. (1995). The
J. Neurophysiol. 31: 659–669.
Nature of Insight. Cambridge, MA: MIT Press.
Streletz, L. J., Belevich, J. K. S., Jones, S. M., Bhusan, A.,
Sterr, A., Muller, M. M., Elbert, T., Rockstroh, B., Pantev,
Shah, S. H., and Herbison, G. (1995). Transcranial mag-
C., and Taub, E. (1998a). Perceptual correlates of changes
netic stimulation: Cortical motor maps in acute spinal cord
in cortical representation of fingers in blind multifinger
injury. Brain Topogr. 7(3): 245–250.
Braille readers. J. Neurosci. 18: 4417–4423.
Stryker, M. (1991). Temporal associations. Nature 354:
Sterr, A., Muller, M. M., Elbert, T., Rockstroh, B., Pantev,
108–109.
C., and Taub, E. (1998b). Changed perceptions in Braille
readers. Nature 391: 134–135.
References 435
Stryker, M. P., and Sherk, H. (1975). Modification of Tallal, P., Miller, S., and Fitch, R. H. (1993). Neuro-
cortical orientation selectivity in the cat by restricted visual biological basis of speech: A case for the preeminence of
experience: A reexamination. Science 190: 904–906. temporal processing. Ann. N. Y. Acad. Sci. 682: 27–47.
Stryker, M. P., Sherk, H., Leventhal, A. G., and Hirsch, Tanaka, J. W., and Farah, M. J. (1993). Parts and wholes in
H. V. (1978). Physiological consequences for the cat’s visual face recognition. Q. J. Exp. Psychol. 46A: 225–245.
cortex of e¤ectively restricting early visual experience with
Tanaka, K. (1994). Inferotemporal cortex and object vision.
oriented contours. J. Neurophysiol. 41: 896–909.
Annu. Rev. Neurosci. 19: 101–139.
Sugita, Y. (1996). Global plasticity in adult visual cortex
Tanaka, K., Saito, H., Fukada, Y., and Moriya, M. (1991).
following reversal of visual input. Nature 380: 523–526.
Coding visual images of objects in the inferotemporal
Sundareswaran, V., and Vaina, L. (1994). Learning direc- cortex of the macaque monkey. J. Neurophysiol. 66: 170–
tion in global motion: Two classes of psychophysically- 189.
motivated models. In G. Tesauro, D. Touretzky, and T.
Tanaka, Y., and Sagi, D. (1998). Long-lasting, long-range
Leen, eds., Advances in Neural Information Processing Systems,
detection facilitation. Vision Res. 38: 2591–2599.
vol. 7, pp. 917–924, San Francisco: Morgan Kaufmann.
Tanne, D., and Sagi, D. (1995). Visual learning can be
Sutherland, N. S., and Macintosh, N. J. (1971). Mechanisms
reversed by task. Invest. Ophthalmol. Vis. Sci. 36: 376.
of Animal Discrimination Learning. New York: Academic
Press. Tarr, M., and Bültho¤, H. H. (1995). Is human object rec-
ognition better described by geon structural descriptions or
Sutherland, N. S., Mackintosh, N. J., and Mackintosh, J.
by multiple views? J. Exp. Psychol. Hum. Percept. Perform.
(1963). Simultaneous discrimination training of Octopus
21: 1494–1505.
and transfer of discrimination along a continuum. J. Comp.
Physiol. Psychol. 56: 150–156. Tarr, M., and Pinker, S. (1989). Mental rotation and ori-
entation dependence in shape recognition. Cogn. Psychol.
Suzuki, W. A., and Amaral, D. G. (1994). Topographic
21: 233–282.
organization of the reciprocal connections between the
monkey entorhinal cortex and the perirhinal and para- Taub, E. (1994). Overcoming learned nonuse: a new be-
hippocampal cortices. J. Neurosci. 14: 1856–1877. havioral medicine approach to physical medicine. In J. G.
Carlson, S. R. Seifert, and N. Birbaumer, eds., Clinical
Suzuki, W. A., and Amaral, D. G. (1995). Perirhinal and
Applied Psychophysiology, 185–220. New York: Plenum.
parahippocampal cortices of the macaque monkey: Corti-
cal a¤erents. J. Comp. Neurol. 349: 1–36. Taub, E., Crago, J. E., and Uswatte, G. (1998). Constraint-
induced movement therapy: A new approach to treatment
Swindale, N. V. (1981). Absence of ocular dominance
in physical rehabilitation. Rehab. Psychol. 43: 152–170.
patches in dark-reared cats. Nature 290: 332–333.
Taub, E., Miller, N. E., Novack, T. A., Cook, E. W.,
Szentágothai, J. (1973). Synaptology of the visual cortex.
Fleming, W. C., Nepomuceno, C. S., Connell, J. S., and
In R. Jung (ed.), Handbook of Sensory Physiology. Vol. 7,
Crago, J. E. (1993). Technique to improve chronic motor
Central Visual Information, 269–324. Berlin: Springer.
deficit after stroke. Arch. Phys. Med. Rehab. 74: 347–354.
Talairach, J., and Tournoux, P. (1988). Co-planar Stereo-
Taub, E., Uswatte, G., and Elbert, T. (forthcoming). Stroke
taxic Atlas of the Human Brain. Stuttgart: Thieme.
rehabilitation and the functional significance of cortical
Tallal, P., Miller, S. L., Bedi, G., Byma, G., Wang, reorganization for behavior and perception. Am. Psychol.
X., Nagarajan, S. S., Schreiner, C., Jenkins, W. M., and
Thorek, K., and Sinha, P. (2001). Qualitative representa-
Merzenich, M. M. (1996). Language comprehension in
tions for recognition. In preparation.
language-learning impaired children improved with acous-
tically modified speech. Science 271: 81–84. Tighe, T. J., and Tighe, L. S. (1968). Perceptual learning
in the discrimination processes of children: An analysis of
436 References
five variables in perceptual pretraining. J. Exp. Psychol. 77: Topka, H., Cohen, L. G., Cole, R. A., and Hallett, M.
125–134. (1991). Reorganization of corticospinal pathways follow-
ing spinal cord injury. Neurology 41: 1276–1283.
Tinazzi, M., Zanette, G., Polo, A., Volpato, D., Man-
ganotti, P., Bonato, C., Testoni, R., and Fiaschi, A. (1997). Torkkola, K. (1996). Blind separation of convolved sources
Transient dea¤erentation in humans induces rapid modu- based on information maximization. In Proceedings of the
lation of primary sensory cortex not associated with sub- IEEE Workshop on Neural Networks and Signal Processing,
cortical changes: A somatosensory evoked potential study. Kyoto, Japan, September.
Neurosci. Lett. 223: 21–24.
Tovee, M. J. (1995). Face recognition: What are faces for?
Tolhurst, D. J., and Barfield, L. P. (1978). Interactions Curr. Biol. 5: 480–482.
between spatial frequency channels. Vision Res. 18: 951–
Tovee, M. J., Rolls, E. T., and Ramachandran, V. S.
958.
(1996). Rapid visual learning in neurones of the primate
Tolhurst, D. J., and Heeger, D. J. (1997). Contrast- temporal visual cortex. NeuroReport 7: 2757–2760.
normalization and threshold models of the responses of
Tovee, M. J., Rolls, E. T., Trevis, A., and Bellis, R. P.
simple cells in cat striate cortex. Vis. Neurosci. 14(2): 293–
(1993). Information encoding and the responses of single
309.
neurons in the primate temporal visual cortex. J. Neuro-
Tolman, E. C. (1948). Cognitive maps in rats and men. physiol. 70: 640–654.
Psych. Rev. 55: 189–208.
Traversa, R., Cicinelli, P., Bassi, A., Rossini, P. M., and
Tomblin, J. B., and Quinn, M. A. (1983). The contribu- Bernardi, G. (1997). Mapping motor cortical reorganiza-
tion of perceptual learning to performance on the repeti- tion after stroke. Stroke 28: 110–117.
tion task. J. Speech Hear. Res. 26: 369–372.
Treisman A. (1996). The binding problem. Curr. Opin.
Tommerdahl, M., Delemos, K. A., Favorov, O. V., Metz, Neurobiol. 6: 171–178.
C. B., Vierck, C. J., Jr., and Whitsel, B. L. (1998).
Treisman, A., and Gelade, G. A. (1980). A feature inte-
Response of anterior parietal cortex to di¤erent modes
gration theory of attention. Cogn. Psychol. 12: 97–136.
of same-site skin stimulation. J. Neurophysiol. 80: 3272–
3283. Treisman, A., Vieira, A., and Hayes, A. (1992). Automa-
ticity and preattentive processing. Am. J. Psychol. 105: 341–
Tommerdahl, M., Delemos, K. A., Vierck, C. J., Jr.,
362.
Favorov, O. V., and Whitsel, B. L. (1996). Anterior parietal
cortical response to tactile and skin-heating stimuli applied Trobalon, J. B., Chamizo, V. D., and Mackintosh, N. J.
to the same skin site. J. Neurophysiol. 75: 2662–2670. (1992). Role of context in perceptual learning in maze
discriminations. Q. J. Exp. Psychol. B. 44: 57–73.
Tong, Y. C., Black, R. C., Clark, G. M., Forster, I. C.,
Millar, J. B., O’Loughlin, B. J., and Patrick, J. F. (1979). A Trobalon, J. B., Sansa, J., Chamizo, V. D., and Mackintosh,
preliminary report on a multiple-channel cochlear implant N. J. (1991). Perceptual learning in maze discriminations.
operation. J. Laryngol. Otol. 93(7): 679–695. Q. J. Exp. Psychol. B. 43: 389–402.
Tong, Y. C., Millar, J. B., Clark, G. M., Martin, L. F., Troje, N., and Bültho¤, H. (1996). Face recognition under
Busby, P. A., and Patrick, J. F. (1980). Psychophysical and varying poses: The role of texture and shape. Vision Res.
speech perception studies on two multiple-channel coch- 36: 1761–1771.
lear implant patients. J. Laryngol. Otol. 94: 1241–1256. Troscianko, T., and Harris, J. (1988). Phase discrimination
Tootell, R. B. H., Switkes, E., Silverman, M. S. and in chromatic compound gratings. Vision Res. 28: 1041–
Hamilton, S. L. (1988). Functional anatomy of macaque 1049.
striate cortex—II. Retinotopic organization. J. Neurosci. 8: Troyer, T. W., Krukowski, A. E., Priebe, N. J., and
1531–1568. Miller, K. D. (1998). Contrast-invariant orientation tuning
References 437
in cat visual cortex: Thalamocortical input tuning and Ullman, S. (1990). 3-dimensional object recognition. Cold
correlation-based intracortical connectivity. J. Neurosci. 18: Spring Harb. Symp. Quant. Biol. 55: 889–898.
5908–5927.
Ullman, S. (1995). Sequence seeking and counter streams:
Ts’o, D. Y., and Gilbert, C. D. (1988). The organization A computational model for bidirectional information flow
of chromatic and spatial interactions in the primate striate in the visual cortex. Cereb. Cortex 5: 1–11.
cortex. J. Neurosci. 8: 1712–1727.
Ullman, S., and Basri, R. (1991). Recognition by linear
Ts’o, D. Y., Gilbert, C. D., and Wiesel, T. N. (1986). combinations of models. IEEE Trans. Patt. Anal. Mach.
Relationships between horizontal interactions and func- Intell. 13: 992–1005.
tional architecture in cat striate cortex as revealed by cross-
Ungerleider, L. G. (1995). Functional brain imaging studies
correlation analysis. J. Neurosci. 6: 1160–1170.
of cortical mechanisms for memory. Science 270: 769–775.
Tsotsos, J. K. (1990). Analyzing vision at the complexity
Ungerleider, L. G., Ga¤an, D., and Pelak, V. S. (1989).
level. Behav. Brain Sci. 13: 423–469.
Projections from inferior temporal cortex to prefrontal
Tulving, E., Hayman, C. A., and Macdonald, C. A. (1991). cortex via the uncinate fascicle in rhesus monkeys. Exp.
Long-lasting perceptual priming and semantic learning in Brain Res. 76: 473–484.
amnesia: A case experiment. J. Exp. Psychol. Learn. Mem.
Ungerleider, L., and Mishkin, M. (1982). Two cortical
Cogn. 17: 595–617.
visual systems. In D. Ingle, M. Goodale, and R. Mansfield,
Tulving, E., and Schacter, D. L. (1990). Priming and eds., Analysis of Visual Behavior, 549–586. Cambridge, MA:
human memory systems. Science 247: 301–306. MIT Press.
Turk, M., and Pentland, A. (1991). Eigenfaces for recog- Uttley, A. M. (1959). The design of conditional probability
nition. J. Cogn. Neurosci. 3: 71–86 computers. Inform. Control 2: 1–24.
Turnure, C. (1972). Perceptual learning in young children: Vaina, L. M., Belliveau J. W., des Roziers, E. B., and
Varied context and stimulus labels. Am. J. Psychol. 85: 339– Ze‰ro, T. (1998). Neural systems underlying learning
349. and representation of global motion. Proc. Natl. Acad. Sci.
U S A 95: 12657–12662.
Turrigiano, G. G., Leslie, K. R., Desai, N. S., Rutherford,
L. C., and Nelson, S. B. (1998). Activity-dependent scal- Valentine, T. (1988). Upside-down faces: A review of the
ing of quantal amplitude in neocortical neurons. Nature e¤ect of inversion upon face recognition. Br. J. Psychol. 79:
391: 892–895. 471–491.
Tversky, B., and Hemenway, K. (1984). Objects, parts and Valentine, T. (1991). Representation and process in face
categories. J. Exp. Psychol. Gen. 113: 169–193. recognition. In R. Watt, ed., Vision and Visual Dysfunction,
vol. 14, pp. 107–124. London: Macmillan.
Uhl, F., Kretschmer, T., Lindinger, G., Goldenberg, G.,
Lang, W., Oder, W., and Deecke, L. (1994). Tactile mental Valverde, F. (1986). Intrinsic neocortical organization:
imagery in sighted persons and in patients su¤ering from Some comparative aspects. Neurosci. 18: 1–23.
blindness early in life. Electroenc. Clin. Neurophysiol. 91:
van Ee, R. (2001). Perceptual learning without feedback
249–255.
and the stability of stereoscopic slant estimation. Perception
Uhl, F., Podreka, I., Steiner, M., and Deecke, L. (1993). 30: 95–114
Increased regional cerebral blood flow in inferior occipital
Van Essen, D. C., Felleman, D. J., DeYoe, E. A., Olavarria,
cortex and cerebellum of early blind humans. Neurosci.
J., and Knierim, J. (1990). Modular and hierarchical orga-
Lett. 150: 162–164.
nization of extrastriate visual cortex in the macaque mon-
Ullman, S. (1979). The Interpretation of Visual Motion. key. Cold Spring Harb. Symp. Quant. Biol. 55: 679–696
Cambridge, MA: MIT Press.
438 References
van Hateren, J. H. (1992). A theory of maximizing sensory Volchan, E., and Gilbert, C. D. (1994). Interocular transfer
information. Biol. Cybern. 68: 23–29. of receptive field expansion in cat visual cortex. Vision Res.
35: 1–6.
van Hateren, J. H., Ruderman, D. L. (1998). Independent
component analysis of natural image sequences yields Volkmann, A. (1858). Über den Einfluss der Übung auf
spatio-temporal filters similar to simple cells in primary das Erkennen räumlicher Distanzen. Berichte über die Ver-
visual cortex. Proc. R. Soc. Lond. B 265: 2315–2320. handlungen der Sächsischen Gesellschaft der Wissenschaft zu
Leipzig, mathmatische und physische Abtheilung 10: 38–69.
Van Hoesen, G. W., and Pandya, D. N. (1975). Some con-
nections of the entorhinal (area 28) and perirhinal (area 35) Volkmann, A. W. (1863). Physiologische Untersuchungen im
cortices of the rhesus monkey: 1. Temporal lobe a¤erents. Gebiete der Optik. Leipzig: Breitkopf und Härtel.
Brain Res. 95: 1–24.
Von Bonin, G., and Bailey, P. (1947). The Neocortex of
van Leeuwen, C. (1990). Perceptual-learning systems as Macaca mulatta. Urbana: University of Illinois Press.
conservative structures: Is economy an attractor? Psychol.
Von der Heydt, R., and Peterhans, E. (1989). Mechanisms
Res. 52: 145–152.
of contour perception in monkey visual cortex—I. Lines
Van Nes, F. L., and Bouman, M. A. (1967). Spatial mod- of pattern discontinuity. J. Neurosci. 9: 1731–1748.
ulation transfer in the human eye. J. Opt. Soc. Am. 57:
von der Malsburg, C. (1987). In The Neural and Molecular
401–406.
Basis of Learning: Dahlem Konferenzen, 1987, ed. J. P.
Vapnik, V. (1995). The Nature of Statistical Learning Theory. Changeux and M. Konishi, 411–432. New York: Wiley.
Berlin: Springer.
von der Malsburg, C. (1999). The what and why of bind-
Vasama, J.-P., Mäkelä, J. P., Pyykkö, I., and Hari, R. ing: The modeler’s perspective. Neuron 24: 95–104.
(1995). Abrupt unilateral deafness modifies function of
von Noorden, G. K. (1990). Binocular Vision and Ocular
human auditory pathways. NeuroReport 6: 961–964.
Motility: Theory and Management of Strabismus. St. Louis:
Vetter, T., and Poggio, T. (1994). Symmetric 3D objects Mosby.
are an easy case for 2D object recognition. Spat. Vis. 8(4):
Vuilleumier, P., and Sagiv, N. (2001). Two eyes make a
443–453.
pair: Facial organization and perceptual learning reduce
Vetter, T., and Poggio, T. A. (1996). Image synthesis from visual extinction. Neuropsychologia 39: 1144–1149.
a single example view. In Computer Vision—ECCV ’96:
Wachsmuth, E., Oram, M. W., and Perrett, D. I. (1994).
Notes in Computer Science. Cambridge, U.K.: Springer.
Recognition of objects and their component parts:
Vidyasagar, T. R., and Stuart, G. W. (1993). Perceptual Responses of single units in the temporal cortex of the
learning in seeing form from motion. Proc. R. Soc. Lond. B macaque. Cereb. Cortex 4: 509–522.
Biol. Sci. 254: 241–244.
Wahba, G. (1979). Convergence rates of ‘‘thin plate’’
Vlek, C. A., and Werner, H. H. (1973). Learning relative smoothing splines when the data are noisy. In T. Gasser
frequency distributions: Some perceptual and cognitive and M. Rosenblatt, eds., Smoothing Techniques for Curve
factors. J. Exp. Psychol. 100: 106–115. Estimation, 233–245. Berlin: Springer.
Vogels, R., and Orban, G. A. (1985). The e¤ect of practice Walk, R. D. (1978). Perceptual learning. In E. C. Car-
on the oblique e¤ect in line orientation judgements. Vision terette and M. P. Friedman, eds., Handbook of Perception,
Res. 25: 1679–1687. vol. 9, pp. 257–298. New York: Academic Press.
Vogels, R., and Orban, G. A. (1994). Does practice in Walker, S., Bruce, V., and O’Malley, C. (1995). Facial
orientation discrimination lead to changes in the response identity and facial speech processing: Familiar faces and
properties of macaque inferior temporal neurons? Eur. J. voices in the McGurk e¤ect. Percept. Psychophys. 57: 1124–
Neurosci. 6: 1680–1690. 1133.
References 439
Wallach, H., and O’Connell, D. N. (1953). The kinetic cortex determined by timing of tactile stimulation. Nature
depth e¤ect. J. Exp. Psychol. 45: 205–217. 378: 71–75.
Wallach, H., O’Connell, D. N., and Neisser, U. (1953). Watanabe, T., Nanez, J. E., and Sasaki, Y. (2001). Per-
The memory e¤ect of visual perception of three-dimen- ceptual learning without perception. Nature 413: 844–
sional form. J. Exp. Psychol. 45: 360–368. 848.
Wallis, G., and Baddeley, R. (1997). Optimal unsupervised Watson, C. S. (1980). Time course of auditory percep-
learning in invariant object recognition. Neural Comput. tual learning. Ann. Otol. Rhinol. Laryngol. Suppl. 89: 96–
9(4): 883–894. 102.
Wallis, G., and Bültho¤, H. (1998). Using a ‘‘virtual illusion’’ Watson, C. S. (1991). Auditory perceptual learning and
to put parallax in its place. Perception 27 ECVP suppl.: 19. the cochlear implant. Am. J. Otol. 12(Suppl.): 73–79.
Wallis, G., and Bültho¤, H. (1999). Learning to recognize Waugh, S. J., Levi, D. M., and Carney, T. (1993). Orien-
objects. Trends Cogn. Sci. 3: 22–31. tation, masking, and vernier acuity for line targets. Vision
Res. 33: 1619–1638.
Wallis, G., and Bültho¤, H. (2001). E¤ects of temporal
association on recognition memory. Proc. Natl. Acad. Sci. Webster, M. A., De Valois, K. K., and Switkes, E. (1990).
U S A 98: 4800–4804. Orientation and spatial-frequency discrimination for lumi-
nance and chromatic gratings. J. Opt. Soc. Am. A7: 1034–
Wallis, G., and Rolls, E. T. (1997). A model of invariant
1049.
object recognition in the visual system. Prog. Neurobiol. 51:
167–194. Wehrhahn, C., and Rapf, D. (2001). Perceptual learning
of apparent motion mediated through ON- and OFF-
Walsh, V., Ashbridge, E., and Cowey, A. (1998). Cortical
pathways in human vision. Vision Res. 41: 353–358.
plasticity in perceptual learning demonstrated by trans-
cranial magnetic stimulation. Neuropsychologia 36: 363– Weiller, C., Chollet, F., Friston, K. J., Wise, R. J. S., and
367. Frackowiak, R. S. J. (1992). Functional reorganization of
the brain in recovery from striatocapsular infarction in
Walsh, V., and Booth, M. (1997). Perceptual learning:
man. Ann. Neurol. 31: 463–472.
Insight in sight. Curr. Biol. 7: R249–251.
Weinberger, N. M. (1995). Dynamic regulation of recep-
Waltz, D. (1975). Generating semantic descriptions from
tive fields and maps in the adult sensory cortex. Annu. Rev.
drawings of scenes with shadows. In P. Winston, ed., The
Neurosci. 18: 129–158.
Psychology of Computer Vision. New York: McGraw-Hill.
Weinberger, N. M., and Bakin, J. S. (1998). Learning-
Waltz, D. L. (1972). Generating semantic descriptions
induced physiological memory in adult primary auditory
from drawings of scenes with shadows. In The Psychology of
cortex: Receptive field plasticity, model, and mechanisms.
Computer Vision, ed. P. H. Winston New York: McGraw-
Audiol. Neurootol. 3: 145–167.
Hill.
Weiss, I. (1988). Projective invariants of shapes. In Pro-
Wang, G., Tanaka, K., and Tanifuji, M. (1996). Optical
ceedings of the IEEE Conference on Computer Vision and Pat-
imaging of functional organization in the monkey infero-
tern Recognition, pp. 291–297.
temporal cortex. Science 272: 1665–1668.
Weiss, T., Miltner, W. H., Adler, T., Bruckner, L., and
Wang, G., Tanifuji, M., and Tanaka, K. (1998). Functional
Taub, E. (1999). Decrease in phantom limb pain associated
architecture in monkey inferotemporal cortex revealed by
with prosthesis-induced increased use of an amputation
in vivo optical imaging. Neurosci. Res. 32: 33–46.
stump in humans. Neurosci. Lett. 272: 131–134.
Wang, X., Merzenich, M. M., Sameshima, K., and Jenkins,
Weiss, Y., Edelman, S., and Fahle, M. (1993). Models of
W. M. (1995). Remodeling of hand representation in adult
perceptual learning in vernier hyperacuity. Neural Comput.
5: 695–718.
440 References
Wertheimer, M. (1938). Laws of Organization in Perceptual Wills, A. J., and McLaren, I. P. (1998). Perceptual learn-
Forms. London: Harcourt, Brace, Jovanovich. ing and free classification. Q. J. Exp. Psychol. B. 51: 235–
270.
Westheimer, G. (1976). Di¤raction theory and visual
hyperacuity. Am. J. Optom. Physiol. Opt. 53: 362–364. Willshaw, D. J., Buneman, O. P., and Longuet-Higgins,
H. C. (1969). Non-holographic associative memory. Nature
Westheimer, G. (1979). Cooperative neural processes
222: 960–962.
involved in stereoscopic acuity. Exp. Brain Res. 36: 585–
597. Wilson, B. S., Finley, C. C., Lawson, D. T., Wolford,
R. D., Eddington, D. K., and Rabinowitz, W. M. (1991).
Westheimer, G. (2001). Is peripheral visual acuity suscep-
Better speech recognition with cochlear implants. Nature
tible to perceptual learning in the adult? Vision Res. 41:
352: 236–238.
47–52.
Wilson, H. R. (1980). A transducer function for threshold
Westheimer, G., Shimamura, K., and McKee, S. P. (1976).
and suprathreshold human vision. Biol. Cybern. 38: 171–
Interference with line orientation sensitivity. J. Opt. Soc.
178.
Am. 66: 332–338.
Wilson, H. R., and Humanski, R. (1993). Spatial frequency
Whitaker, D., and McGraw, P. V. (2000). Long-term visual
adaptation and contrast gain control. Vision Res. 33: 1133–
experience recalibrates human orientation perception. Nat.
1149.
Neurosci. 3: 13.
Wilson, H. R., and Wilkinson, F. (1997). Evolving con-
Widrow, B., and Stearns, S. D. (1985). Adaptive Signal
cepts of spatial channels in vision: From independence to
Processing. Englewood Cli¤s, NJ: Prentice Hall.
nonlinear interactions. Perception 26: 939–960.
Wiesel, T. N. (1982). Postnatal development of the visual
Wilson, M. A., and McNaughton, B. L. (1993). Dynamics
cortex and the influence of environment. Nature 299: 583–
of the hippocampal ensemble code for space. Science 261:
591.
1055–1058.
Wiesel, T. N., and Hubel, D. H. (1963). Single-cell
Wilson, P., and Snow, P. J. (1987). Reorganization of the
responses in striate cortex of kittens deprived of vision in
receptive fields of spinocervical tract neurons following
one eye. J. Neurophysiol. 26: 1003–1017.
denervation of a single digit in the cat. J. Neurophysiol. 57:
Wiesel, T. N., and Hubel, D. H. (1965). Comparison of 803–818.
the e¤ects of unilateral and bilateral eye closure on cortical
Winston, P., ed. (1975). The Psychology of Computer Vision.
unit responses in kittens. J. Neurophysiol. 28: 1029–1040.
New York: McGraw-Hill.
Wiesel, T. N., and Hubel, D. H. (1965). Binocular inter-
Wiskott, L. (1998). Learning invariance manifolds. In L.
action in striate cortex of kittens reared with artificial
Niklasson, M. Bodén, and T. Ziemke, eds., Proceedings of
squint. J. Neurophysiol. 28: 1041–1059.
the International Conference on Artificial Neural Networks:
Wiggs, C. L., and Martin, A. (1998). Properties and Perspectives in Neural Computing, 555–560. Berlin: Springer.
mechanisms of perceptual priming. Curr. Opin. Neurobiol.
Witkin, A. P. (1981). Recovering surface shape and ori-
8: 227–233.
entation from texture. Artif. Intell. 17(1–3): 17–45.
Williamson, J. R. (1997). A constructive, incremental-
Wohlwill, J. F. (1966). Perceptual learning. Annu. Rev.
learning network for mixture modeling and classification.
Psychol. 17: 201–232.
Neural Comput. 9: 1517–1543.
Wolf, S. L., Lecraw, D. E., Barton, L. A., and Jahn,
Williamson, J. R. (1999). Learning vernier discrimination
B. B. (1989). Forced use of hemiplegic upper extremities
with a model of hierarchical map formation. Invest. Oph-
to reverse the e¤ect of learned nonuse among chronic
thalmol. Vis. Sci. 3976.
References 441
stroke and head injured patients. Exp. Neurol. 104: 125– Xing, J., and Gerstein, G. L. (1996). Networks with lateral
132. connectivity: 3. Plasticity and reorganization of somato-
sensory cortex. J. Neurophysiol. 75: 217–232.
Wolford, G., and Kim, H.-Y. (1992). The role of visible
persistence in backward masking. In From Learning Processes Xu, J., and Wall, J. T. (1997). Rapid changes in brainstem
to Cognitive Processes: Essays in Honor of William K. Estes, ed. maps of adult primates after peripheral injury. Brain Res.
A. F. Healy, S. M. Kosslyn, and R. M. Shi¤rin, vol. 2, pp. 774: 211–215.
161–180. Hillsdale, NJ: Erlbaum.
Xu, J., and Wall, J. T. (1999). Evidence for brainstem and
Wolford, G., Marchak, F., and Hughes, H. (1988). Practice supra-brainstem contributions to rapid cortical plasticity in
e¤ects in backward masking. J. Exp. Psychol. Hum. Percept. adult monkeys. J. Neurosci. 19: 7578–7590.
Perform. 14: 101–112.
Yaginuma, S., Osawa, Y., Yamaguchi, K., and Iwai, D.
Woods, T. M., Cusick, C. G., Pons, T. P., Taub, E., and (1993). Di¤erential functions of central and peripheral
Jones, E. G. (2000). Progressive transneuronal changes in visual field representations in monkey prestriate cortex. In
the brainstem and thalamus after long-term dorsal rhizot- Brain Mechanisms of Perception and Memory: From Neuron to
omies in adult macaque monkeys. J. Neurosci. 20: 3884– Behavior, ed. T. Ono et al., 1–33. New York: Oxford
3899. University Press.
Wright, B. A., Buonomano, D. V., Mahncke, H. W., and Yakovlev, V., Fusi, S., Berman, E., and Zohary, E. (1998).
Merzenich, M. M. (1997). Learning and generalization of Inter-trial neuronal activity in inferior temporal cortex: A
auditory temporal-interval discrimination in humans. J. putative vehicle to generate long-term visual associations.
Neurosci. 17: 3956–3963. Nat. Neurosci. 1: 310–317.
Wright, R. L., and Whittlesea, B. W. (1998). Implicit Yamashita, H. (1993). Perceptual-motor learning in am-
learning of complex structures: Active adaptation and se- nesic patients with medial temporal lobe lesions. Percept.
lective processing in acquisition and application. Mem. Mot. Skills 77: 1311–1314.
Cognit. 26: 402–420.
Yarbus, A. L. (1957). The perception of an image fixed
Wülfing, E. A. (1892). Ueber den kleinsten Gesichtswin- with respect to the retina. Biophysics 2: 683–690.
kel. Zeitschrift für Biologie 29(11): 199–202.
Yin, R. K. (1969). Looking at upside-down faces. J. Exp.
Wurtz, R., Goldberg, M., and Robinson, D. L. (1982). Psychol. Gen. 81: 141–145.
Brain mechanisms of visual attention. Sci. Am. 246(6): 100–
Young, A. W., Hay, D. C., and Ellis, A. W. (1985). The
107.
faces that launched a thousand slips: Everyday errors and
Wurtz, R. P., Vorbruggen, J. C., and von der Malsburg, di‰culties in recognising people. Br. J. Psychol. 76: 495–523.
C. (1990). A transputer system for the recognition of
Young, A. W., Hay, D. C., McWeeny, K. H., Flude,
human faces by labeled graph matching. In R. Eckmiller,
B. M., and Ellis, A. W. (1985). Matching familiar and un-
G. Hartmann, and G. Hauske, eds., Parallel Processing in
familiar faces on internal and external features. Perception
Neural Systems and Computers, 37–41. Amsterdam: Elsevier.
14: 737–746.
Xerri, C., Merzenich, M. M., Jenkins, W., and Santucci, S.
Young, A. W., Hellawell, D., and Hay, D. C. (1987).
(1999). Representational plasticity in cortical area 3b par-
Configural information in face perception. Perception 16:
alleling tactual-motor skill acquisition in adult monkeys.
747–759.
Cereb. Cortex 9: 264–276.
Young, A. W., Rowland, D., Calder, A. J., Etco¤, N.,
Xerri, C., Stern, J. M., and Merzenich, M. M. (1994).
Seth, A., and Perrett, D. I. (1996). Facial expression mega-
Alterations of the cortical representation of the rat ventrum
induced by nursing behavior. J. Neurosci. 14: 1710–1721.
442 References
mix: Tests of dimensional and categorical accounts of ation of cortical and thalamic hindpaw representations by
emotion recognition. Cognition 63: 271–313. modification of walking in adult rats. Soc. Neurosci. Abstr.
22: 1055.
Young, D., Lawlor, P. A., Leone, P., Dragunow, M., and
During, M. J. (1999). Environmental enrichment inhibits Zernicki, B. (1991). Visual discrimination learning in bin-
spontaneous apoptosis, prevents seizures and is neuro- ocularly deprived cats: 20 years of studies in the Nencki
protective. Nat. Med. 5: 448–453. Institute. Brain Res. Rev. 16: 1–13.
Young, M. (1992). Objective analysis of the topological Zernicki, B. (1999). Visual discrimination learning under
organization of the primate cortical visual system. Nature switching procedure in visually deprived cats. Behav. Brain
358: 152–155. Res. 100: 237–244.
Young, M. P., and Yamane, S. (1992). Sparse population Ziemann, U., Corwell, B., and Cohen, L. G. (1998).
coding of faces in the inferotemporal cortex. Science 256: Modulation of plasticity in human motor cortex after fore-
1327–1331. arm ischemic nerve block. J. Neurosci. 18(3): 1115–1123.
Yuille, A. (1991). Deformable templates for face recogni- Zihl, J., and von Cramon, D. (1979). Restitution of visual
tion. J. Cogn. Neurosci. 3(1): 59–71. function in patients with cerebral blindness. J. Neurol.
Neurosurg. Psychiatry 42: 312–322.
Yuille, A. L. (1987). Shape from shading, occlusion and
texture. AI Memo 885. MIT AI Laboratory. Zihl, J., and von Cramon, D. (1985). Visual field recovery
from scotoma in patients with postgeniculate damage.
Yukie, M., Takeuchi, H., Hasegawa, Y., and Iwai, E.
Brain 108: 335–365.
(1990). Di¤erential connectivity of inferotemporal area TE
with the amygdala and the hippocampus in the monkey. In Zipser, K., Lamme, V. A. F., and Schiller, P. H. (1996).
Vision, Memory and the Temporal Lobe, ed. E. Iwai and M. Contextual modulation in primary visual cortex. J. Neurosci.
Mishkin, 129–135. New York: Elsevier. 16: 7376–7389.
Zanker, J. M. (1999). Perceptual learning in primary and Zohary, E., Celebrini, S., Britten, K., and Newsome,
secondary motion vision. Vision Res. 39: 1293–1304. W. T. (1994). Neuronal plasticity that underlies improve-
ment in perceptual performance. Science 263: 1289–1292.
Zeki, S. (1993). A Vision of the Brain. Oxford: Blackwell.
Zohary, E., and Newsome, W. T. (1994). Perceptual
Zelniker, T., and Oppenheimer, L. (1976). E¤ect of dif-
learning in a direction discrimination task is not based
ferent training methods on perceptual learning in impul-
upon enhanced neuronal sensitivity in the STS. Invest.
sive children. Child Dev. 47: 492–497.
Ophthalmol. Vis. Sci. 35: 1663.
Zenger, B., and Sagi, D. (1996). Isolating excitatory and
inhibitory nonlinear spatial interactions involved in con-
trast detection. Vision Res. 36: 2497–2513.
Zenger-Landolt, B., and Fahle, M. (2001). Discriminating
contrast discontinuities: Asymmetries, dipper functions,
and perceptual learning. Vision Res. 41: 3009–3021.
Zepka, R. F., Godde, B., and Dinse, H. R. (2000). Syn-
chronous and asynchronous tactile coactivation controls
distance between stimulated skin representations in soma-
tosensory cortex of adult rats: An optical imaging study.
Soc. Neurosci. Abstr. 26: 933.
Zepka, R. F., Jürgens, M., and Dinse, H. R. (1996). Dif-
ferential time course of use-dependent plastic reorganiz-
Contributors
University of Liverpool
Liverpool, U.K.
Keiji Tanaka
RIKEN Brain Research Institute
Saitama, Japan
Guy Wallis
School of Human Movement Studies
University of Queensland
Brisbane, Australia
Barbara Zenger
Department of Psychology
Stanford University
Stanford, California
This page intentionally left blank
Index
Abrupt learning, 237, 244, 250. See also Insight Attention, xv, 216–217, 253–255
vs. gradual learning, xiv–xv, 250 performance always requires, 271
specific to retinal size, 237–240 selective, 226–227, 253–254, 265
as triggered by any ‘‘easy’’ stimulus, 244–247 as necessary for learning, 261, 263–264
Adaptation, ix as su‰cient for learning, 264–265
of inputs in somatosensory system (see also Plasticity) spatial, 226–227, 254
imaging studies on impact of modified use and practice, Attentional constraints, 271–272
28–29 Attentional hierarchy, 255
role of input statistics, 29–34 Attitudes, personal
therapeutic consequences, 34–35 and early perception, 278
Adaptational changes, coding, 39–42 Auditory cortex, 85–86
Adaptational processes Auditory learning, xiii–xiv
interaction with information processing, 39, 40 Auditory system
scheme of, 40 cortical reorganization in, 142
Adaptive filtering, 120 types of plasticity in, 154 (see also under Cochlear implant)
Aftere¤ects, ix Axial selectivity, 8
Age, e¤ects of, 35
on receptive fields, 27 BCM (Bienenstock, Cooper, and Munro) learning rule,
on temporal sequence representation, 38 363–364
Age dependence of perceptual learning, xvii Berkeley, George, 302
Agnosia, 123 Blind persons, 139. See also Braille readers
Ambiguous figures, 278 role of primary visual areas in, 130
Amnesic patients, xvi–xvii Blood oxygenation level dependent (BOLD) technique,
perceptual learning, 209, 210 131
Amputation, upper limb, 134–135 Braille readers, alterations of somatosensory representation
consecutive alterations in motor representations after, in, 128–129
132, 134 Braille reading, 139
Angular changes in pose, 329 neural networks for, 130, 138
Anterior inferotemporal (AIT) cortex, 220 Brain-derived neurotrophic factor (BDNF), 61, 62
Apperceptive agnosia, 123 Burton’s model of learning new faces, 332–334
Association
of action and stimulus, 122–123 Callosal connections, development and plasticity of, 13, 15
elementary law of, 110 Categorical e¤ects and categorical perception, 329–330,
mechanism of, 251n1 344
vs. tuning, 119–120 Categorization, 96
Associative agnosia, 123 levels of, 319
Associative learning, 251 Classification, ix–x, 348–349, 368–369
and nonassociative learning, viii Clustering algorithms, 344
448 Index
Hyperacuity, visual, 197, 217. See also Vernier acuity in problem solving and perception, 249
computational mechanisms of, 198–200 and rote learning, 236
neuronal mechanisms underlying, 201–202 Interactive activation and competition. See IAC
as probe of perceptual learning, 197–198 Internal criteria and knowledge/information, 373–375
what it is good for, 200–201 Intracortical connections, long-range, 4. See also Callosal
Hyperacuity tasks, 274, 367 connections
training and transfer between, 212–214 axial selectivity, 8
development, 10–15
IAC (interactive activation and competition), 332, 333 patch development, 11–14
IAC with Learning (IACL), 333, 334 patchy connections and orientation columns during,
Ideal edges, 294–295 11
‘‘Ideal’’ output, 295–296 experience-dependent selection, 14
Identifying responses, 96 extent and laminar specificity, 4–5
Illusions, 303, 304 functions of horizontal, 8–10
Illusory contours (ICs), 237, 238 layout, 4–10
Image interpretation, high-level influences on, 276–278 modular selectivity, 5, 8
Immediate input interpretation (III), 230 patchy nature of, 5, 6
Implicit learning, perceptual learning as form of, viii plasticity
Independent component analysis (ICA), 355–361, 363– in the adult, 16–18
365 during development, 11–14
Inferotemporal cortex, anterior, 220 underlying mechanism and substrate, 17–18
Inferotemporal (IT) lobe, 309, 313 types of neurons involved in, 5
cortex of lateral, 97 Intracortical microstimulation (ICMS), 29–30
Inferotemporal (IT) neurons, 97, 112, 114, 116, 123, 312. Invariance, position, 229–231
See also TE cells Invariance learning, temporal continuity as cue to, 313, 315
adult, with responses similar to infants’ responses, 118 Inversion e¤ect, 320
changeability and selectivity in the adult, 77–81
experience-based modification, 114 Kinetic depth e¤ect (KDE), 283, 284, 286, 287
features and representations, 97–99 Köhler, W., 301
biological relevance, 105
complexity, selectivity, generality, and reduction, 99– Learning dynamics and enabling, 268–269
105 Learning-induced plasticity, 19. See also Plasticity, use-
space and distributed representations, 105–107 dependent
highly selective, 102, 103 ‘‘Learning maps,’’ 264, 265
‘‘Infomax’’ principle, 361 Learning (perceptual). See also specific topics
Inhibitory e¤ects, learning, 191 as active process, 248
Injuries, repetitive strain, 34 along a continuum, 267–268
Injury-related invasion, 134 and conflicting stimuli, 259–260
Input conditions, hemisphere-specific, 134 cortical localization of neuronal changes underlying,
Input-output probability density, learning joint, 349–350 206–207
Input statistics defined, vii
and adaptation, 29–34 forms of, 236
vs. attention, 22 goals, 338–341
Insight, ix, 235–236, 247–251, 373 high-level, 275 (see also under Perception)
abrupt improvement through (see Abrupt learning) computational model for incorporating it in early
Hebb’s behavioral criterion for, 243–244 perception, 290–296
Index 451
Positron emission tomography (PET), 129–130 object, 70, 230 (see also Face recognition; Three-
Practice, ix–x dimensional [3-D] object recognition)
Practice e¤ects, 183–185 based on two-dimensional image features, 308–309
Pragmatism in perception, 303–305 object recognition paradigms, 305–309
‘‘Preattentive’’ process, 255 of objects in images without relying on sophisticated
Priming, viii, ix perceptual constructs, 298
repetition, 120 philosophical beginnings, 300–301
visual, 120 Reduction method, 103
‘‘Primitive’’ features, 108–109 di‰culties with, 103, 104
Principal component (PC) representations of images, 332 Redundancy reduction, 364, 365
Principal components analysis (PCA), 326, 328, 357– Regional cerebral blood flow (rCBF), 131
359 Regression, 346–349
Probability density functions (PDFs), 358 Reinforcement, and early perception, 278–279
Projective invariants, 306–307 Repetitive strain injuries, 34
Prosopagnosia, 312 Repetitive transcranial magnetic stimulation (rTMS), 132
Protein kinase C (PKC) substrates, phosphorylation of, Repetitive visual stimulation, and receptive fields in adult
110 striate cortex, 43–47
Prototype face representations, 328–329 Representational theories of vision, 347–348
Punishment, and early perception, 278–279 Representations (cortical), 121
development of, 108–109
Radial basis function (RBF), 345, 347 as dynamic time-based constructs, 134
Receptive fields (RFs), 30–32, 35, 37, 38, 98. See also learning as forming new, 345–347
Visual field Retinal images, organization of, 219–220
in adult striate cortex, repetitive visual stimulation and, Retinal lesions
43–47 acute, 48
artificial scotomas, acute retinal lesions, and, 48 chronic changes in receptive field size and cortical
dea¤erentation and, 47–51 topography induced by, 50–51
Gaussian, 340 cortical immunohistochemistry after chronic, 51
mechanisms for modifying visual, 58, 60–65 topographical reorganization following, 49
LTP-like e¤ects, learning, and training, 62–65 Retinal size, abrupt learning specific to, 237–240
methods of assessing, 36 Retinal stabilized images. See Scotomas, artificial
orientation-specific, 201, 202 Retinotopy, 219
plasticity, 43 e¤ect of dea¤erentation on, 48–51
growth factors and morphological correlates of, 61–62 Reverse hierarchy theory, 265–267, 271
properties, changing as result of perceptual learning, xiv schematic illustration, 266
self-organization, 361 Reward and punishment, and early perception, 278–
Receptive fields (RF) size, 39, 228 279
and anatomical level of visual hierarchy, 220 ‘‘Rodent on a runway,’’ 337
changes in, 39, 61, 63–65, 201 Rote learning, 236, 251n1
chronic, induced by lesions, 50–51, 54–56, 63
and cortical topography, 53–56 Sampling theorem, 199, 200
fast lesion- and training-induced, near lesions in cat Satiation, ix
visual cortex, 56–59 S.B., case of, 301–303, 305
visual stimulation and, 57–58 Scotomas, 17–18
Recognition (visual), 96 artificial, 43
disorders of, 123 and receptive fields, 48
454 Index
Ventroanterior TE, 82
Vernier acuity, 201, 202, 264. See also Hyperacuity
improvement through training, 202–204
Vernier discrimination tasks, 200–204, 206–212, 367, 368,
372, 374
Vernier discrimination thresholds, 206–210
Vernier o¤sets, 200
Vernier stimuli, training with subthreshold, 214, 215
Vibrotactile stimulation, xiii
View specificity, training to override, 312
Vision. See also specific topics
processes underlying, 96
Visual agnosia, 123
Visual areas, primary
role in blind persons, 130
Visual cortex, plasticity of, vii–viii, 107–109. See also
Plasticity
in early visual areas, 109–110
in higher visual areas, 110–116
multiple memory systems, 119–121
time course of learning, 116–119
Visual encounters, consequences of all, 95