The Best Time to Acquire New Skills: Age-related Differences in Implicit Sequence Learning across Human Life Span

Serial reaction time task and the development of implicit learning

In our study, we used a modified version of the Serial Reaction Time (SRT) Task, which is one of most commonly used methods for measuring implicit skill learning. Serial Reaction Time Task is a four-choice reaction time task containing a hidden repeating sequence that the subject comes to predict and learn implicitly (Nissen & Bullemer, 1987; Poldrack et al., 2005). In an SRT study, Meulemans et al. (1998) found that 6- and 10-year-old children showed similar degrees of learning as young adults. In contrast, Thomas et al. (2004) found that the learning performance of young adults was better than 7- to 10-year-old children. Studies investigating implicit skill learning at older ages also revealed inconsistent results. For example, several studies have demonstrated that, for simple repeating patterns (in the SRT task), the extent of implicit sequence learning in elderly adults was comparable to young adults (Frensch & Miner, 1994; D. V. Howard & Howard, 1992; D. V. Howard & Howard, 1989). Moreover, in a recent study, Gaillard et al. (2009) found that young (22-year-old), middle-aged (45-year-old), and elderly (71-year-old) participants performed at the same level.

The studies mentioned above used fixed (deterministic) sequences, which can be easily learned, making it less possible to detect age-related differences in learning. Furthermore, they cannot purely determine the acquired sequence-specific knowledge because these tasks (finger-tapping, classical SRT) confound general improvements with sequence-specific learning. Here, we used a modified version of the SRT task, the Alternating Serial Reaction Time (ASRT) task (J. H. Howard, Jr. & Howard, 1997), which enabled us to measure the “pure” sequence-specific learning distinguished from general improvements. In the classical SRT task, the structure of a sequence is deterministic with the stimuli following a simple cyclically repeating pattern (e.g., 213412134121341213412…, where numbers refer to distinct events within the repeating 21341 pattern). By contrast, in the ASRT task (J. H. Howard, Jr. & Howard, 1997; Remillard, 2008), repeating events alternate with random ones. Thus, the location of every second stimulus on the screen was determined randomly. If, for instance, the sequence was 12341234…, where the numbers represent locations on the screen, the sequence of stimuli would be 1R2R3R4R1R2R3R4R… in the ASRT paradigm, with R representing a random element. Therefore, the location of every second stimulus on the screen was determined randomly. Because fixed, sequence-specific and random stimuli were alternating, some sequences of three events (called ‘triplets’) occurred more frequently than others. For example, in the above illustration 1×2, 2×3, 3×4 and 4×1 would occur often, whereas 1×3 or 4×2 would occur infrequently. Following previous studies, we referred to the former as high-frequency triplets and the latter as low-frequency triplets (Nemeth, Janacsek, Londe, et al., 2010; Song, Howard, & Howard, 2007). Previous studies have shown that as people practice the ASRT task, they respond more quickly to the high than low frequency triplets, revealing probabilistic, sequence-specific learning (J. H. Howard, Jr. & Howard, 1997; Song et al., 2007). This learning is statistical in nature because it depends on the frequency of the event sequences. Thus, the RT difference between the high and low frequency triplets in this ASRT task is a measure of human sensitivity to the relative raw probabilities of events observed implicitly in their environment (Perruchet & Pacton, 2006). In addition, the participants are not generally aware of the alternating structure of the sequences, even after extended practice, or when sensitive recognition tests are used to assess explicit knowledge (D. V. Howard, et al., 2004; J. H. Howard, Jr. & Howard, 1997; Song et al., 2007). Thus, the ASRT task is more implicit than the classical deterministic sequence learning tasks.

Using the ASRT task, recent studies have shown that, although elderly adults can also learn the higher-order structure of these complex sequences, they showed age-related deficits (D. V. Howard et al., 2004; J. H. Howard, Jr. & Howard, 1997; Nemeth, Janacsek, Londe et al., 2010). Both young and elderly adults were able to learn third-order dependencies (1RR2RR3RR1RR2RR3RR…) although the elderly participants performed at a lower level than the younger participants (Bennett, Howard, & Howard, 2007). Whereas several studies investigated implicit learning in children using the ASRT task (Barnes, et al., 2008; Barnes, Howard, Howard, Kenealy, & Vaidya, 2010; Nemeth, Janacsek, Balogh, et al., 2010), no child-adult comparison of implicit skill learning performance has yet been reported.

In summary, previous studies have addressed the development and aging in implicit skill learning, but no studies have examined age-related differences from childhood to old age with identical methods. Furthermore, in contrast to general skill improvements, using a probabilistic sequence learning task (ASRT) can help us to reveal the age-related differences of the underlying mechanisms of complex skill learning by measuring explicitly the sensitivity to raw probabilities of high and low frequency events. Therefore, in this study, we compared the implicit probabilistic sequence learning across the age range of 4–85 years.

Method

Results

Overall RT’s significantly differed among the age groups (main effect of AGE: F(8,412)=107.11, p<0.001, η_p²=0.675; Table 1). The RT decreased significantly between each group from 4–6 to 18–29 years of age (all p’s<0.04), they were similar between the age groups of 18–29 and 30–44 (p>0.38) and significantly increased after 44 years of age (p’s<0.001) (Figure 1a). The accuracy monotonically increased over the years (main effect of AGE: F(8,412)=16.94, p<0.001, η_p²=0.25) (Figure 1b).

The comparison of RT to high vs. low probability triplets showed a surprising pattern of implicit sequence learning across the age groups. Even though there was a significant learning at all ages because the RT’s were faster for high frequency than to low frequency triplets (main effect of TRIPLET TYPE: F(1,412)=333.7, p<0.001, η_p²=0.45, all p’s <0.03) (Figure 1c), the magnitude of this difference was not uniform. Although the age groups differed significantly from each other in sequence learning (TRIPLET TYPE × AGE interaction: F(8,412)=6.79, p<0.001, η_p²=0.12), the post-hoc test revealed that learning was significantly higher in the 4- to 12-year-old groups than in any other group in the 14–85 range (p’s<0.02). There was no difference in learning between the 14–59 years of age (p’s>0.37), whereas the magnitude of learning decreased significantly in the 60–85-year-old group (p’s<0.02). Thus, learning high probability events was uniformly effective until the age of 12 where it reduced significantly and remained at a lower level of sensitivity until the age of 60 (Figure 1c).

However, it is a long-standing issue in developmental and aging studies how to compare groups with different baseline speeds. A customary approach to this problem is to analyze the data using z-transformation. Therefore, we calculated the z-scores within each subject (thus, each participant's own mean and SD was used to transform that participant's data (see, for example, Christ, White, Mandernach, & Keys, 2001) and conducted an ANOVA based on these z-scores (Figure 2). ANOVA revealed significant sequence-specific learning (main effect of TRIPLET TYPE: F(1,412)=320.12, p<0.001, η_p²=0.44), but the extent of learning differed across groups (TRIPLET TYPE×AGE interaction: F(8,412)=8.91, p<0.001, η_p²=0.15). We found that the participants from 9 years of age showed similar extent of sequence learning as the adult groups to 44 years of age (all p’s>0.25), but the learning in 4–8 years of age was smaller compared to these adult groups (all p’s<0.014). At older ages, there was a decline in the sequence learning, with both the 45–59 and 60–85-year-old group differing significantly from the groups between the 11–44 years of age (all p’s<0.025).

The analysis of the response accuracy further enhanced the picture emerging from the results with RTs. We found a significantly greater accuracy for high- than low-frequency triplets (main effect of TRIPLET TYPE: F(1,412)=217.14, p<0.001, η_p²=0.345). Although all age groups older than 6 showed significant sequence learning (all p’s<0.011), the age groups differed significantly in the strength of the sequence learning (TRIPLET TYPE × AGE interaction: F(8,412)=3.73, p<0.001, η_p²=0.07). Whereas groups between 7 and 44 years of age showed similar degrees of learning, this was significantly higher than the youngest (4–6) and the two oldest (45–59 and 60–85) groups (p’s<0.03) (Figure 1d).

Figure 3 shows the individual data for sequence learning measured by raw RT, accuracy and z-transformed RT data. The pattern of this data is consistent with the ANOVA results: 1) children between 4–12 years of age showed greater sequence learning as measured by raw RT, whereas 2) in sequence-specific learning as measured by accuracy and z-transformed RTs adults exhibited the highest performance.

Discussion

The goal of the present study was to investigate the differences in implicit skill learning across the human life span. This work extends previous studies (Bennett et al., 2007; Gaillard et al., 2009; J. H. Howard, Jr. & Howard, 1997; Meulemans et al., 1998; Thomas et al., 2004) in two ways: 1) it examined a wide range of ages between 4–85 years, and 2) it used a probabilistic task, which enabled us to measure the “pure” sequence-specific learning defined by the sensitivity to raw probabilities of high and low frequency events. We found that the 4- to 12-year-old age groups showed the strongest learning effect measured by the raw RT difference scores. Around the age of 12, we found a striking transition to less pronounced sequence-specific learning, as measured by smaller differences between the responses to high and low frequency triplets. Confirming earlier results (D. V. Howard et al., 2004; J. H. Howard, Jr. & Howard, 1997; Nemeth, Janacsek, Londe et al., 2010), we found that this learning capacity was significantly reduced in the oldest age group. Thus, in contrast to the developmental invariance (Reber, 1993) and the age-related changes approaches (Meulemans et al., 1998; Vinter & Perruchet, 2000), our results demonstrate a gradual decline in learning across the lifespan.

Sequence learning scores based on the accuracy and raw reaction time showed different curves: the former one is a bell-shaped curve, whereas the latter is a gradually declining curve (Figure 1c–d). Hence, these two types of learning scores can reflect different underlying mechanisms and brain systems. The accuracy learning score may be more related to attention, mainly voluntary attention, whereas the RT learning score may be related to involuntary attention and intuitive processes (Burgess, Gilbert, & Dumontheil, 2007; Prinzmetal, McCool, & Park, 2005). The relatively weaker accuracy learning effects in children and older groups may be due to the underdeveloped/deteriorating attentional brain circuits connected to the frontal lobe.

Our study raised a methodological issue which affects the developmental studies in general. It is a long-standing issue in the literature how to compare groups with different baseline speeds. It could be argued that the youngest groups in our study have larger raw RT learning scores (more difference between high and low frequency triplets) because they have more room to do so, given their longer RTs. However this argument does not seem to hold in our study, because the oldest group showed equally long RTs as 9–10 year-olds, yet they had a four-fold reduced learning score (Figure 1). Along a similar line of argument, the accuracy measures of the youngest group (4–6 year-olds) and the oldest groups (45–85 year-olds) showed a marked difference (~8%) in general accuracy, giving more room for the young group to produce larger differences between high and low frequency triplets, yet the actual sequence-learning (i.e. high-low differences in accuracy) showed no difference. In contrast, the two smallest age groups had almost identical general accuracy yet there was a more than two-fold increase in sequence learning in the second youngest group (7–8 year-olds). Thus, our data suggests no linear relationship between the general magnitude of reaction time/accuracy and the learning measures.

A second customary approach to the problem of comparing groups with different baseline speeds is to analyze the data using z-transformation. Z-scores of our results show a different picture than raw the RT data analysis: the learning performances from 9 to 44 years of age are similar with weaker performance in the younger and older age groups. The z-transformation is often used to control general processing speed across different age groups in developmental studies. However, the main function of z-transformation is not to control the processing speed, but to normalize the distribution of responses. Thus, z-transformation has mathematical assumptions about the form of distributions and is therefore not theory independent (Yap, Balota, Sibley, & Ratcliff, in press). Z scoring fully adjusts for processing speed only if all participants have the same type of distributions. Therefore using z-scores in developmental studies might be misleading. Furthermore, it is unclear how general processing speed and variability contributes to learning and performance in different ages. In recent years, several studies analyzed the variability and noise across a wide age-range (Der & Deary, 2006; Hultsch, MacDonald, Hunter, Levy-Bencheton, & Strauss, 2000; Rabbitt, Osman, Moore, & Stollery, 2001). For example, Rabbitt et al. (2001) found that people’s fastest RT’s were relatively unaffected by age, but the number of unnecessarily slow responses was higher in older ages, and, thus, the increase in the mean RT was a result of increasing variability, which was an important component of cognitive aging. Moreover, several studies outline that the noise and the performance variability enables adaptive plasticity of motor skills (Slifkin & Newell, 1998, 1999; Turner & Brainard, 2007) and high variability can support effective learning and performance (Sanger, 2010). Thus, based on previous studies and on our analyses, we suggest that both the processing speed (mean reaction time) and variability are inherent aspects of development and aging. We think that the z-transformation eliminates these inherent aspects of learning, making the transformed results difficult to understand and explain.

Based on the reasoning above, we propose that, the raw RT difference between the high and low frequency triplets in the ASRT task is a measure of human sensitivity to the relative raw probabilities of events observed implicitly in their environment. Thus, our results show a marked decrease in this sensitivity around the age of 12, which is in contrast to the traditional view of a steady improvement of cognitive learning abilities until late in adulthood (Craik & Bialystok, 2006). However, this discrepancy might be explained based on a shift in the structural development of implicit learning. Although the raw probabilities of the sensory environment are important for learning and both infants (Aslin, Saffran, & Newport, 1998; Fiser & Aslin, 2002; Saffran et al., 1996; Saffran, Johnson, Aslin, & Newport, 1999) and adults (Fiser & Aslin, 2001; Hunt & Aslin, 2001) are highly sensitive to them, there is an ongoing debate on how using these simple probabilities can lead to highly complex knowledge of the world, such as sensory invariances and development of a language (Gomez & Gerken, 1999; Marcus, Vijayan, Rao, & Vishton, 1999). Recent studies proposed that using an internally stored structured model of the world that emerges based on past experience together with probabilistic learning could help to address this issue and also provide evidence that humans might implement such a strategy during implicit learning (Orban, Fiser, Aslin, & Lengyel, 2008; Tenenbaum, Kemp, Griffiths, & Goodman, 2011). In this framework, as the internal model develops with experiences becoming more influential, internal interpretations of events become more elaborate and less directly related to their raw probabilities. A recent study argued that from a normative standpoint, existence of multiple learning mechanisms in the brain (cf. model-free vs. model-based learning) with an uncertainty–based arbitration between them would be computationally optimal (Daw et al., 2005). Anchoring this hypothesis biologically, the presumed mechanisms related to these two types of learning were suggested to be related to the prefrontal areas and temporal lobe of the cortex, respectively (Daw et al., 2005). Support for the separated and complementary nature of the prefrontal- and medial temporal lobe (MTL)-dependent learning based on internal models vs. basal ganglia-dependent model-free learning comes from various studies investigating learning under specific conditions. These studies showed that obstructing the PFC and/or MTL by a demanding secondary task (Foerde, Knowlton, & Poldrack, 2006) do not adversely affect implicit learning. Other studies found that inserting a task between the learning sessions (Brown & Robertson, 2007a, 2007b), performing a working memory and an implicit learning task simultaneously (Filoteo, Lauritzen, & Maddox, 2010), or a neuropharmacological blockage (Frank, O'Reilly, & Curran, 2006) even had a positive effect on performance in implicit learning task. Importantly, it is known that the cortical areas connected to the internal models related to model-based learning become truly functional late in the development, around age of 12 (Blakemore & Choudhury, 2006; Giedd, et al., 1999), which is about the age when we found the sudden decrement in sensitivity to the relative raw probabilities. We propose, that this enhanced functionality signals the shift when the system adapts efficiently to more complex aspects of the world by relying more on internal model-based interpretations, while somewhat neglecting the raw probabilities of the sensory input (Figure 4a–b), and therefore, decreasing the ability to develop and stabilize fundamentally new basic competences. Thus the seemingly paradoxical result of gradually becoming less sensitive to basic statistics, if timed appropriately, could be the optimal strategy for human implicit learning in general.

Our results did not reveal any differences between the young adults and middle-age groups. Salthouse’s (1996) “simultaneity mechanism” theory of cognitive aging predicts the age-related deficits in probabilistic sequence learning (Curran, 1997; J. H. Howard, Jr. & Howard, 1997). Feeney, Howard & Howard (2002) found age-related deficits in pattern sensitivity in “older” (mean age: 49.4) compared with “younger” (mean age: 41.4) middle-aged groups. These different results could be related to that Feeney used a smaller sample size, a longer version of the ASRT and different method of analysis.

What are the underlying mechanisms of the decreased performance of the elderly group? In a recent fMRI study, Dennis and Cabeza (in press) showed that older adults recruited the MTL for implicit learning, and this activation was significantly greater, while striatal activity decreased in older people compared with young adults during implicit learning. In a recent study, Rieckmann, Fisher & Backman (2010) found similar results: in young adults during the learning session, the activation of the striatum increased, but the that of MTL decreased. By contrast, in older adults, sequence learning positively related to activation increases in both the striatum and MTL. Using multimodal imaging measures, Giorgo et al. (2010) found extensive reductions in the gray matter volume in aging, but reductions were detected earlier in the frontal cortex. Furthermore, a recent diffusion tensor imaging aging study by Bennett et al. (in press) found that the caudate–dorsolateral prefrontal cortex (DLPFC) and hippocampus-DLPFC tract integrity were related to ASRT sequence learning. The caudate-DLPFC tract integrity decreased in the older ages, mediating age-related differences in sequence learning. Within the computational framework proposed by Daw and collegues (2005), these findings can be interpreted as a deterioration in three mechanisms that contribute to the age-related decline in skill learning: 1) reduced detection of probabilities, 2) rigidity of internal models and/or 3) more restricted connections between internal models and probability detection (Figure 4c). Thus, not only the model-free, but also the model-based learning, might be limited in older ages. Future studies are needed to systematically examine the underlying neural mechanisms of age-related differences in skill learning.

In summary, based on our raw RT results we suggest that acquiring fundamentally new skills that cannot be derived from skills already possessed is the most effective before adolescence, and it might be largely based on the fronto-striatal circuitry, such as the basal ganglia and cerebellum, in agreement with earlier skill learning models (Doyon et al., 2009; Hikosaka et al., 1999; Hikosaka et al., 2002) and computational learning models (Daw et al., 2005). Our findings are in good agreement with everyday life experience showing that an early (~ before 12 years) start of learning some sports, music instruments, second language, etc. often leads to higher level of competence. These results may have implications for the development of learning and memory, facilitating new skill training and pedagogic methods (e.g., for teaching languages) and may also contribute to the understanding of neurodevelopmental and age-related disorders (e.g., autism, SLI, dyslexia and dementia) and lead to relevant treatment options.