PDF Intelligence Science and Big Data Engineering Visual Data Engineering 9Th International Conference Iscide 2019 Nanjing China October 17 20 2019 Proceedings Part I Zhen Cui Ebook Full Chapter
PDF Intelligence Science and Big Data Engineering Visual Data Engineering 9Th International Conference Iscide 2019 Nanjing China October 17 20 2019 Proceedings Part I Zhen Cui Ebook Full Chapter
PDF Intelligence Science and Big Data Engineering Visual Data Engineering 9Th International Conference Iscide 2019 Nanjing China October 17 20 2019 Proceedings Part I Zhen Cui Ebook Full Chapter
https://textbookfull.com/product/intelligence-science-and-big-
data-engineering-8th-international-conference-
iscide-2018-lanzhou-china-august-18-19-2018-revised-selected-
papers-yuxin-peng/
https://textbookfull.com/product/intelligent-data-engineering-
and-automated-learning-ideal-2019-20th-international-conference-
manchester-uk-november-14-16-2019-proceedings-part-i-hujun-yin/
https://textbookfull.com/product/advanced-hybrid-information-
processing-third-eai-international-conference-adhip-2019-nanjing-
china-september-21-22-2019-proceedings-part-i-guan-gui/
https://textbookfull.com/product/advances-in-knowledge-discovery-
and-data-mining-23rd-pacific-asia-conference-pakdd-2019-macau-
china-april-14-17-2019-proceedings-part-i-qiang-yang/
https://textbookfull.com/product/knowledge-science-engineering-
and-management-12th-international-conference-ksem-2019-athens-
greece-august-28-30-2019-proceedings-part-i-christos-douligeris/
https://textbookfull.com/product/intelligent-data-engineering-
and-automated-learning-ideal-2019-20th-international-conference-
manchester-uk-november-14-16-2019-proceedings-part-ii-hujun-yin/
Zhen Cui · Jinshan Pan · Shanshan Zhang ·
Liang Xiao · Jian Yang (Eds.)
Intelligence Science
LNCS 11935
Founding Editors
Gerhard Goos
Karlsruhe Institute of Technology, Karlsruhe, Germany
Juris Hartmanis
Cornell University, Ithaca, NY, USA
Intelligence Science
and Big Data Engineering
Visual Data Engineering
9th International Conference, IScIDE 2019
Nanjing, China, October 17–20, 2019
Proceedings, Part I
123
Editors
Zhen Cui Jinshan Pan
Nanjing University of Science Nanjing University of Science
and Technology and Technology
Nanjing, China Nanjing, China
Shanshan Zhang Liang Xiao
Nanjing University of Science Nanjing University of Science
and Technology and Technology
Nanjing, China Nanjing, China
Jian Yang
Nanjing University of Science
and Technology
Nanjing, China
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
General Chairs
Lei Xu Shanghai Jiao Tong University, China
Xinbo Gao Xidian University, China
Jian Yang Nanjing University of Science and Technology, China
Program Chairs
Huafu Chen University of Electronic Science and Technology
of China, China
Zhouchen Lin Peking University, China
Kun Zhang Carnegie Mellon University, USA
Zhen Cui Nanjing University of Science and Technology, China
Organization Chairs
Liang Xiao Nanjing University of Science and Technology, China
Chen Gong Nanjing University of Science and Technology, China
Publication Chairs
Shanshan Zhang Nanjing University of Science and Technology, China
Wankou Yang Southeast University, China
Program Committee
Mingming Gong The University of Melbourne, Australia
Joseph Ramsey Carnegie Mellon University, USA
Biwei Huang Carnegie Mellon University, USA
Daniel Malinsky Johns Hopkins University, USA
Ruben Sanchez-Romero Rutgers University, USA
Shohei Shimizu RIKEN, Japan
Ruichu Cai Guangdong University of Technology, China
Shuigeng Zhou Fudan University, China
Changdong Wang Sun Yat-sen University, China
Tao Lei Shaanxi University of Science & Technology, China
viii Organization
Sparse Unmixing for Hyperspectral Image with Nonlocal Low-Rank Prior . . . 506
Feiyang Wu, Yuhui Zheng, and Le Sun
3Q: A 3-Layer Semantic Analysis Model for Question Suite Reduction . . . . . 219
Wei Dai, Siyuan Sheni, and Tieke Hei
Lei Xu1,2(B)
1
Centre for Cognitive Machines and Computational Health (CMaCH), SEIEE,
Shanghai Jiao Tong University, Minhang, Shanghai, China
[email protected]
2
Neural Computation Research Centre, Brain and Intelligence Sci-Tech Institute,
ZhangJiang National Lab, Shanghai, China
http://www.cs.sjtu.edu.cn/~lxu/
Abstract. Deep bidirectional Intelligence (BI) via YIng YAng (IA) sys-
tem, or shortly Deep IA-BI, is featured by circling A-mapping and I-
mapping (or shortly AI circling) that sequentially performs each of five
actions. A basic foundation of IA-BI is bidirectional learning that makes
the cascading of A-mapping and I-mapping (shortly A-I cascading)
approximate an identical mapping, with a nature of layered, topology-
preserved, and modularised development. One exemplar is Lmser that
improves autoencoder by incremental bidirectional layered development
of cognition, featured by two dual natures DPN and DCW. Two typical
IA-BI scenarios are further addressed. One considers bidirectional cogni-
tion and image thinking, together with a proposal that combines theories
of Hubel-Wiesel’s versus Chen’s. The other considers bidirectional inte-
gration of cognition, knowledge accumulation, and abstract thinking for
improving implementation of searching, optimising, and reasoning. Par-
ticularly, an IA-DSM scheme is proposed for solving a doubly stochastic
matrix (DSM) featured combinatorial tasks such as travelling salesman
problem, and also a Subtree driven reasoning scheme is proposed for
improving production rule based reasoning. In addition, some remarks
are made on relations of Deep IA-BI to Hubel and Wiesel theory, Sperry
theory, and A5 problem solving paradigm.
one of ancient Chinese yIng yAng logo1 . First, real bodies or patterns in the
Actual world or shortly A-domain (the domain that is visible or named yAng)
are mapped along the inward direction into the inner coding domain or shortly
I-domain (the domain that is invisible or named yIng). This mapping transfers
real body to information coding seed like a yAng or male animal (named A-
mapping), performing Abstraction tasks such as perception, recognition, and
cognition. Second, codes, concepts or symbols in the inner I-domain are mapped
along the outward direction named I-mapping (a mapping from coding seed to
real body in A-domain, i.e., acting like a yIng or female animal) to make Inference
tasks that may be categorised into image thinking and abstract thinking.
The A-mapping and then I-mapping circling (or shortly A-I circling) performs
each of five actions sequentially. The first action is acquiring data X and features
that describe X. As to be further addressed in Sect. 2, the process is featured
by either deep neural networks (NN) or convolutional NN (CNN) that proceeds
layer by layer to perform hierarchical extraction and abstraction. The 1981 Nobel
prize winners Hubel and Wiesel [24,25] developed a theory that explains how this
proceeds in a manner of hierarchy as illustrated in Fig. 1(a). This H-W theory
has greatly impacted the subsequent efforts in the studies of artificial intelligence
and neural networks, including the recent more than decade long bloom of deep
learning driven AI studies.
The second action performs abstraction by an inner code Y that indicates one
among labels or concepts, allocates chances among candidates, gets a dimension-
Fig. 1. Deep IA-BI (i.e., bidirectional Intelligence via YIng YAng system)
1
“Ying” is spelled “Yin” in the current Chinese Pin Yin system that could be back-
tracked to over 400 years from the initiatives by M. Ricci and N. Trigault. But, the
length of ‘Yin’ lost its harmony with Yang, thus ‘Ying’ is preferred since 1995 [42].
Deep IA-BI and Five Actions in Circling 3
ally much reduced code as an inner representation of X, and even forms an icon
or a subtree structure. This action not only performs perception and recognition,
but also provides cognitions and evidences to the third action Inner-memory that
accumulates knowledge and evidence.
The knowledge and evidence come from two sources. One acquires, mem-
orizes, and organises various knowledges via education, e.g., in a formulation
of knowledge graph. The other not only gets evidences to these organised and
structured knowledges, but also adds in concepts and cognitions from the second
action.
The fourth action inference may be jointly activated by the status of its
previous two actions and possibly some short-cut signals as well, which performs
either or both of following two manners:
Image Thinking. It is also closely related to what called concrete thinking
elsewhere, and is usually referred to thought process based on dependencies
among either or both of real/concrete things and their mappings/images in the
I-domain of brain. The key point is specific, concrete, and detailed. Cognition
via A-mapping trained by supervised deep learning is just one scenario. Another
scenario happens when there is no teaching label. Whether the perceived Y by
the A-mapping X → Y makes a sense is verified by checking if X̂ generated
by the I-mapping Y → X̂ approximates X as closely as possible. Based on the
cognition by X → Y , the I-mapping Y → X̂ performs various tasks that map
X of input patterns into Z of simplified patterns, enriched patterns, and trans-
formed patterns, as well as imaginary and creative patterns. Typical examples
include language to language, text to image, text to sketch, sketch to image,
image to image, 2D image to 3D image, past to future, image to caption, image
to sentence, music to dance, ..., etc. All these image thinking tasks are performed
by an A-I cascading and featured by an information flow that varies layer by
layer in a topological preservation manner, as if displayed by an image sequence.
Other details are referred to Sect. V.A in Ref. [57].
Abstract Thinking. It is also closely related to what called rational thinking,
and is usually referred to thought process based on either or both of causal rela-
tions among events and logical relations among concepts in a broad, general and
non-specific sense, typically described by symbolic or graphical representations.
Typical examples are searching, selecting, optimising, reasoning, and planning,
which are performed in a discrete space of individual or combinatoric choices.
Traditionally, abstract thinking is performed by I-mapping that searches among
discrete space according to knowledges and evidences accumulated and organised
in the third action, which usually encounters intractable computing difficulties.
Exemplified by AlphaGo [38] and AlphaGoZero [39], searching performances can
be significantly improved with help of one appropriate A-mapping via deep neu-
ral networks that provides either or both of probabilistic selecting polices and
heuristic values.
Following either or both of image thinking and abstract thinking, the fifth
action is implementation of communication (verbal, writing, gesturing, postur-
ing, etc.) and control (motoring, monitoring, steering, etc.) as desired.
4 L. Xu
but differs in that the encoder and decoder are overlapped, resulting in several
favourable characteristics, see further details in Table 3 of [57].
First, neurons per layer in the encoder and the decoder are bidirectionally
connected pairwisely to the corresponding neurons on the corresponding layer,
with each neuron taking a dual role both in the encoder and decoder as illus-
trated in Fig. 2(d), which is referred as Duality in Paired Neurons (DPN) [57].
Specifically, this duality leads to the following three extensions of autoencoder.
(a) Fast-lane Lmser: as illustrated in Fig. 2(b), neurons per layer in the encoder
are directed pairwisely to their counteractions in the decoder, in a role similar
to skip connections in U-net [35], ResNet [16] and DenseNet [22].
(b) Feedback Lmser: as illustrated in Fig. 2(c), neurons per layer in the decoder
are directed pairwisely to their counteractions in the encoder, in a role similar
to those in recurrent neural networks (RNN) for enhancing robustness.
(c) Lmser and flexible Lmser: as illustrated in Fig. 2(d), each neuron per layer j
enforces the activity v (j) in the encoder and the activity u(j) in the decoder
to become identical, which implies that the encoding action from the j-th
layer up to the top and the decoding action from the top down to the j-
th layer jointly perform an identical mapping v (j) = u(j) . In other words,
in addition to seeking one global identical mapping X = X̂ as AE does,
identical mappings are also sought for distributed implementation of the PC
principle, not only from the j-th layer to the top and then down to the j-th
layer, but also between the j-th layer and the j + 1-th layer.
Second, bidirectional layered development of cognition is also considered in
LMSER by another dual nature called Duality in Connection Weights (DCW),
with same connecting weights between every two consecutive layers taking a
Fig. 2. Lmser differs from Autoencoder. (a) Autoencoder without the dualities DPN
and DCW, (b) Fast-lane Lmser with DPN only in skip direction, (c) Feedback Lmser
with DPN only in feedback direction, (d) Lmser with DPN and DCW.
6 L. Xu
dual role both in the encoder and in the decoder. From the j-th to the j + 1-th
layer, we have Aj = WjT in Lmser as illustrated in Fig. 2(d) while Aj is learned
without the constraint Aj = WjT as illustrated in Fig. 2(e). When Wj is an
orthogonal matrix, Aj = WjT approximately acts as its pseudo-inverse such that
Aj WjT = Wj WjT . In other words, DCW enhances distributed implementation of
the PC principle consecutively from the j-th to the j + 1-th layer.
Insights may also come from an incremental bidirectional layered develop-
ment of cognition, as illustrated in Fig. 2(b). Perception and learning start at the
bottom layer, i.e., one layer Lmser that learns templates of feature extraction, as
demonstrated empirically in 1991 [40,46], with details referred to a recent rein-
vestigation [23] and a systematical survey [57]. Then, another layer is topped
on the learned one layer Lmser, and learning is further made on the resulted
two layer Lmser, . . . , so on and so forth. This procedure is similar to learning
stacked RBMs [18,19], as illustrated in Fig. 3(c). The importances of the DPN
and DCW dualities may be interestingly backtracked from Hinton’s progresses
on bidirectional learning from Helmholtz machines [11,17] that also considers an
architecture that locates between AE and Lmser (i.e., with DPN but without
DCW) to stacked RBMs that share both the DPN and DCW dualities.
Started from early fifties of the last century, the main stream studies on percep-
tion and cognition proceed along the direction of feature detector hypothesis on
visual information process, exemplified by the 1981 Nobel Prize winners Hubel
and Wiesel who developed a feature detection theory [24,25]. They found that
some neurons called simple cells that fires rapidly when presented with lines at
one angle, while others responded best to another angle, and also complex cells
that detect edges regardless of where they were placed in the receptive field and
could preferentially detect motion in certain directions. Feature detection pro-
ceeds from detecting direction at the bottom gradually up to organise into more
complicated patterns in a manner of hierarchy, as illustrated in Fig. 1(a).
8 L. Xu
This H-W theory has greatly impacted the subsequent efforts on modelling of
intelligence in the studies of artificial intelligence and neural networks, roughly
summarised into three streams as follows:
(1) The early computing power was very limited in the later seventies and the
early eighties to support the demands of AI studies. Marr and Poggio [30]
proposed a simplified scheme of only three layers, which got popularised in
the AI literature during the eighties of the last century. However, recent
advances of deep learning for computer visions have actually abandoned this
stream.
(2) Fukushima is the first who attempted to build up a computational model
that is loyal to the H-W scheme, proposed Cognitron [13] and then developed
into Neocognitron [14], in which equations for S-cells and C-cells are both
provided with the connections for S-cells modified by learning while ones for
C-cells pre-fixed [15]. Neocognitron may be regarded as a junior version of
the current convolutional neural networks (CNN) [26,27]. The convolutional
layers of CNN share the same point as S-cells that aims at recognising stimu-
lus patterns based on the geometrical similarity (Gestalt), while the pooling
layers of CNN like C-cells. The key difference is that Neocognitron modifies
weights by self-organising while CNN modifies weights by back-propagation.
(3) There are also two roads on how feature detectors or receptive fields are
developed. One considers Gabor functions or other wavelets, with parameters
estimated from data. The other considers how neurons in a simple model
shown in Fig. 4(a) develop receptive fields [2,6,29,36,37,40,46]. At the first
stage, it has been found that receptive fields come from the evolution of
weight vectors by a Hebb-type learning. A typical example is Linsker’s feed-
forward linear network [29] with local connections from a lower layer to its
next layer, which is functionally similar to a convolutional layer. Each neuron
in this linear network is a special case of the one illustrated in Fig. 4(a) at
simply a linear activation s(r) = r. It follows [36,37] that such a type of
evolution may come from either of principal component analysis (PCA) and
maximum information transfer.
Fig. 4. Development of feature detectors or receptive fields. (a) simple neuron that
develops receptive field, (b) feature detector by one layer RBM, (c) Lmser and RBM
share a same structure and also a common term (i.e., two red dashed boxes) in learning.
(Color figure online)
of Gestalt psychology. Proceeding far beyond the notation “whole is more than
the simple sum of actions”, Chen suggests that “holistic registration is priori to
local analyses” and emphasises topological structure in visual perception. This
‘priori’ has two meanings. One implies that global organisations, determined by
topology, are the basis that perception of local geometrical properties depends
on. The other is that topological perception (based on physical connectivity)
occurs earlier than the perception of local geometrical properties. Though there
have been a number of evidences that support this global precedence, there has
no computational approach to successfully implement or even illustrate, which
remains an attractive challenge.
We start with considering abstract thinking in a general sense that deals with
symbolic or graphical representations with help of discrete mathematics, featured
by typical tasks as illustrated in Figs. 6 and 7.
The first task is search vs selection. The simple case is selecting among a finite
number of individuals. The same situation can be found in making a decision
or classifying/allocating among a number of choices. Beyond the simple case,
search is performed by sequential decision to find a path, i.e., making selection
Deep IA-BI and Five Actions in Circling 11
per step as illustrated in Fig. 6(a). Typically, search is finding an optimal path in
a tree or spanning a tree as illustrated on the top of Fig. 6(b), which are typically
encountered in solving travelling salesman problem (TSP) and attributed graph
matching (AGM), as well as alphaGo.
The second task is satisfaction vs optimisation. The target of searching can
be not only one or more nodes but also one or more paths, such that some
specifications or conditions are satisfied. Moreover, we may select an optimal
one or ones among all the satisfactory ones. Traditionally, tree searching usually
encounters intractable computing difficulties. Exemplified by AlphaGo [38] and
AlphaGoZero [39], searching performances can be significantly improved with
help of an appropriate A-mapping via deep neural networks that provides either
or both of selecting probabilities and value heuristics. Furthermore, optimisation
may also be made on other tasks of graph analyses, illustrated in Fig. 6(c).
Here, we sketch one IA scheme for solving a doubly stochastic matrix (DSM)
featured combinatorial task, or shortly IA-DSM optimisation, e.g., the TSP task
illustrated in Fig. 6(c). The solution is featured by the following problem
min E(V ),
V
V = [vij ], i, j = 1, · · · , n, vij = 0, 1, vij = 1, vij = 1, (1)
j i
where vij = 1 indicates going from the ith city to the jth city, and there are only
n ones with vij = 1 that defines a TSP tour. The problem is finding a shortest
TSP tour such that a cost E(V ) is minimised, which is typically NP hard.
Since 1994, a class of neural networks algorithms (named LT algorithms)
[8–10,41,43] have been developed by iteratively solving a mirror of Eq. (1) with
Fig. 6. Bidirectional searching. (a) simple selection and depth first search (DFS), (b)
IA-DSM scheme, (c) travelling salesman problem, (d) other graph analyses.
Another random document with
no related content on Scribd:
be determined.
17‒27. The same subject is treated in xxvi. 1‒19, and this fact
has been urged as an argument for the view that chapter ix. is an
addition to the Chronicler’s work (see Introduction p. xxiii). But it is
also reasonable to suppose that the Chronicler would here give a
register of inhabitants of Jerusalem (which could not be included in
the list of the separate tribes), and such a register would probably
give a survey of the Levitical classes.
for the camp of the children of Levi] i.e. the Temple; but the
phrase, which is derived from Numbers ii. 17, in its original context of
course signifies the Tabernacle of the Mosaic period. Doubtless it is
used with the implication that the institution of the gatekeepers dated
back to that age: compare verse 19 ad fin., and contrast verse 22.
19. over the camp of the Lord, keepers] We might expect the
reference to the Temple or Tabernacle to be continued; but, as
nothing is said in the Pentateuch of “keepers of the entry to the
tabernacle,” probably the entry to the camp, not to the tabernacle, is
meant in the present phrase. With this view agrees the mention of
Phinehas (verse 20), for it apparently was the profanation of the
camp in general, not of the tabernacle, which Phinehas avenged
(Numbers xxv. 6‒8), thus earning a blessing (Numbers xxv. 11‒13).
whom David ... did ordain] The Chronicler attributes to David the
organisation of the priests (xxiv. 3), of the Levites (xxiii. 27; xxiv. 31),
of the singers (xxv. 1 ff.), and of the doorkeepers (in this passage). It
has been thought that this verse is at variance with verses 18, 19,
where the Mosaic origin of the gatekeepers seems to be implied. But
in answer it may be said that the Chronicler is guilty of no
inconsistency in ascribing the origin of the doorkeepers to the
Mosaic period and saying here that David and Samuel “ordained
them in their set office,” for the phrase refers, not to their origin, but
to their organisation. For another suggestion see below on verse 26.
the seer] For the title, xxvi. 28, xxix. 29; 1 Samuel ix. 9; and
compare 2 Chronicles xvi. 7.
24. On the four sides] Fuller details are given in xxvi. 14‒18.
28, 29.
Duties of the Levites.
30.
A Priestly Duty.
31, 32.
Other Levitical Duties.
Chapter X.
1‒12 (= 1 Samuel xxxi. 1‒13).
The Defeat, Death, and Burial of Saul.
1‒12. There are several variations between the text given here
and the text of 1 Samuel, to which attention will be called in the
notes below.
and abuse me] i.e. wreak their cruel will upon me; compare
Judges i. 6.
⁶So Saul died, and his three sons; and all his
house died together.
6. all his house] In Samuel “his armourbearer and all his men.”
The reference may be to Saul’s servants: his family was not
exterminated in this battle.
12. took away] i.e. from the walls of Beth-shan (so Peshitṭa).
asked counsel ... spirit] i.e. of the witch of Endor, 1 Samuel xxviii.
7 ff.
we are thy bone and thy flesh] The phrase is not to be taken
strictly as implying kinship, for only the tribe of Judah could say “The
king is near of kin to us” (2 Samuel xix. 42). The other tribes mean
that they will obey David as though he were their own kin.
2. the Lord thy God said] Compare verses 3, 10; 1 Samuel xvi.
1‒13.