CONNECTIONIST LEARNING: IS IT TIME TO RECONSIDER THE FOUNDATIONS?

[Asim Roy]

Dear Moderator,

Please post this version of the note if you have not posted it
already. I have reformatted it. My sincere apologies to those who
get multiple copies.

Asim Roy
----------------------------------

This note is to summarize the discussion that took place at ICNN'97
(International Conference on Neural Networks) in Houston in June on
the topic of "Connectionist Learning: Is It Time to Reconsider the
Foundations?" ICNN'97 was organized jointly by INNS (International
Neural Network Society) and the IEEE Neural Network Council. The
following persons were on the panel to discuss the questions being
raised about classical connectionist learning:

1. Shunichi Amari
2. Eric Baum
3. Rolf Eckmiller
4. Lee Giles
5. Geoffrey Hinton
6. Teuvo Kohonen
7. Dan Levine
8. Jean Jacques Slotine
9. John Taylor
10. David Waltz
11. Paul Werbos
12. Nicolaos Karayiannis
13. Asim Roy

Nick Karayiannis, General Chair of ICNN'97, moderated the panel
discussion.

Appendix 1 has the issues/questions being raised about classical
connectionist learning. A general summary of the panel discussion
as it relates to those questions is provided below. Appendix 2
provides a brief summary of what was said by individual panel
members. In general, the individual panel members provided their
own summaries. In some cases, they modified my draft of what they
had said. This document took a while to prepare, given the fact
that many of us go on vacation or to conferences during summer.


	A GENERAL SUMMARY OF THE PANEL DISCUSSION

1) On the issue of using memory for learning, many panel members
strongly supported the idea and argued in its favor, saying that
humans indeed store information in order to learn. Although there
was no one who actually opposed the idea of using memory to learn,
some still tend to believe that memoryless learning does indeed
occur in certain situations, such as in the learning of motor
skills. (I can argue very strongly that memory is indeed used in
"every" learning situation, even to acquire motor skills! I
plan to send out a memo on this shortly.)

2) On the question of global/local learning, many panelists agreed
that global learning mechanisms are indeed used in the brain and
pointed out the role of neuromodulators in transferring information
to appropriate parts of the brain. Some others justified global
mechanisms by saying that certain kinds of learning are only
possible with "nonlocal" mechanisms. Again, although there was no
one who vigorously opposed the idea of using global mechanisms to
learn, some thought that some form of local learning may also be
used by the brain in certain situations.

3) On the question of network design, several panelists argued that
the brain must indeed know how to design networks in order to
store/learn new knowledge and information. Some suggested that this
design capability is derived from "experience" (as opposed to
"inheritance" - David Waltz), while others mentioned 
"punishment/reward" mechanisms as its source (John Taylor) or
implied it through the notion of "control of adaptivity" (Teuvo
Kohonen). Shunichi Amari emphasized the network design capability
from a robotics point of view, while Eric Baum said that learning
beyond inherited structures involves knowing how to design
networks. Perhaps all of us agree that we do indeed inherit some 
network structures through evolution/inheritance. But I did not
hear anybody argue that our algorithms should not include network
design as one of its tasks.

		SOME PERSONAL REMARKS ON THIS DEBATE

I have come out of this debate with deep respect for the field and
for many of its highly distinguished and prominent scholars. It has
never been an acrimonious debate. I think most of them had been
very open minded in examining the facts and arguments against
classical connectionist learning. I had vigorous arguments with
some of them, but it was always friendly and very respectful. And I
think we all had fun arguing about these things. I think it bodes
well for the science. The culture of a scientific field depends
very much on its topmost scholars. I couldn't be among a better set
of scholars with higher levels of intellectual integrity.

And to be honest, I was indeed pleasantly surprised when I was
nominated for the INNS Governing Board membership by Prof. Shunichi
Amari. After publicly challenging some of the core connectionist
ideas, I was afraid that I was going to be a permanent outcast
in this field. I hope that will not be true. I hope to be part of
this field.

I think the ICNN'97 debate was very significant and useful. First,
it engaged some of the most distinguished scholars in the field.
Second, there are some very significant statements from many of
these scholars. Paul Werbos was the first to acknowledge that
memory is indeed used for learning. I think that was an important
first step in this debate. But then, there are many others. For
example, Shunichi Amari's call for "a new type of mathematical
theories of neural computation" is very significant indeed. And so
is Teuvo Kohonen's acknowledgment of a "third level of 'control of
synaptic plasticity' that was ignored in the past in
connectionism." And note Dan Levine's statement that "it is indeed
time to reconsider the foundations of connectionist learning,"
despite his emphasis that the work of the last thirty years should
be built upon rather than discarded. And John Taylor's remark that
"classical connectionism perhaps has too narrow a view of the
brain" and that "connectionism should not be limited to traditional
artificial neural networks." And Eric Baum's remarks on the brain
being a multiagent system and on the limitations of classical
connectionist in explaining this multiagent behavior. And Lee
Giles' call for a "deeper foundation for intelligence processing."
And David Waltz's story about the learning experiences of Marvin
Minsky's dog will certainly be a classic. He helped to hammer in
the point strongly that humans do indeed use memory to learn.


		WHAT DID THE DEBATE REALLY ACCOMPLISH?


Overall, the debate has established the following:

1) It is no longer necessary for our learning algorithms to have
local learning laws similar to the ones in back propagation or
perceptron or Hopfield net. This will allow us to develop much more
robust and powerful learning algorithms using means that may be
"nonlocal" in nature. In other words, we should be free to develop
new kinds of algorithms to design and train networks without the
need to use a local learning law.

2) The learning algorithms can now have better access to
information much as humans do. Humans actually have access to all
kinds of information in order to learn. And they use memory to
remember some of it so that they can use it in the thinking and
learning process. The idea of "memoryless" learning in classical
connectionist learning is unduly restrictive and completely
unnatural. There is no biological or behavioral basis for it. So,
our learning algorithms should now be allowed to store learning
examples in order to learn and have access to other kinds of
information. This will allow the algorithms to look at the
information about a problem, understand its complexity and then
design and train an appropriate net.

All this can perhaps be summarized in one sentence: Overall, it
fundamentally changes the nature of algorithms that we might call
"brain-like." So Shunichi Amari's call for "a new type of
mathematical theories of neural computation" couldn't have been
more appropriate.

In my opinion, the debate on connectionist learning does not end
here - it is just the beginning. We should continue to ask critical
questions and engage ourselves in vigorous debate. It doesn't make
sense for a scientific field to work for years on building a theory
that falls apart on first rigorous common sense examination. Many
technological advances depend on this field. So we need to guard
against these major pitfalls.

Perhaps one of the existing newsgroups or a new one can accommodate
such open debates, bringing together neuroscientists, cognitive
scientists and connectionists. I don't think we can be isolated
anymore. The Internet is helpful and allows us to communicate
across disciplines on a worldwide basis. We should no longer be the
lonely researcher with very restricted interactions. I was once a
lonely researcher with many questions in my mind. So I went to
Stanford University during my sabbatical and sat in David
Rumelhart's and Bernie Widrow's classes to ask all kinds of
questions. But there must be a better way to ask such outrageous
questions.

An important issue for the field is that of setting standards for
our algorithms. It is imperative that we define some "external
behavioral characteristics" for our so-called brain-like autonomous
learning algorithms, whatever kind they may be. But I hope that
this is at least a first step towards defining and developing a
more rigorous science. We cannot continue to "babysit" our
so-called learning algorithms. They need to be truly autonomous.

With regards to all,
Asim Roy
Arizona State University


------------------------------------------------------------
			APPENDIX 1

PANEL TITLE:  "Connectionist Learning: Is it Time to Reconsider the
Foundations?"

			ABSTRACT

Classical connectionist learning is based on two key ideas. First,
no training examples are to be stored by the learning algorithm in
its memory (memoryless learning). It can use and perform whatever
computations are needed on any particular training example, but
must forget that example before examining others. The idea is to
obviate the need for large amounts of memory to store a large
number of training examples. The second key idea is that of local
learning - that the nodes of a network are autonomous learners.
Local learning embodies the viewpoint that simple, autonomous
learners, such as the single nodes of a network, can in fact
produce complex behavior in a collective fashion. This second idea,
in its purest form, implies a predefined net being provided to the
algorithm for learning, such as in multilayer perceptrons.

Recently, some questions have been raised about the validity of
these classical ideas. The arguments against classical ideas are
simple and compelling. For example, it is a common fact that humans
do remember and recall information that is provided to them as part
of learning. And the task of learning is considerably easier when
one remembers relevant facts and information than when one doesn't.
Second, strict local learning (e.g. basic back propagation type
learning) is not a feasible idea for any system, biological or
otherwise. It implies predefining a network "by the system" without
having seen a single training example and without having any
knowledge at all of the complexity of the problem. Again, there is
no system that can do that in a meaningful way. The other fallacy
of the local learning idea is that it acknowledges the existence of
a "master" system that provides the design so that autonomous
learners can learn.

Recent work has shown that much better learning algorithms, in
terms of computational properties (e.g. designing and training a
network in polynomial time complexity, etc.), can be developed if
we don't constrain them with the restrictions of classical
learning. It is, therefore, perhaps time to reexamine the ideas of
what we call "brain-like learning."

This panel will attempt to address some of the following questions
on classical connectionist learning:

1.  Should memory  be used for learning? Is memoryless learning an
unnecessary restriction on learning algorithms?
2.  Is local learning a sensible idea? Can better learning
algorithms be developed without this restriction?
3.  Who designs the network inside an autonomous learning system
such as the brain?

---------------------------------------------------------
			APPENDIX 2

BRIEF SUMMARY OF INDIVIDUAL REMARKS


1) DR. SHUNICHI AMARI:

Dr. Amari focused mainly on the neural network design and
modularity of learning.  Classical connectionist learning has
treated microscopic aspects of learning where local generalized
Hebbian rule plays a fundamental role.  However, each neuron works
in a network so that learning signals may be synthesized in the
network nonlocally.  He also said that, based on microscopic local
learning rules, more macroscopic structural learning emerges such
that a number of experts differentiate to play different roles
cooperatively.  This is a basis for concept formation and
symbolization of microscopic neural excitations.  He stresses that
we need a new type of mathematical theories of neural computation.

---------------

Shun-ichi Amari is a Professor-Emeritus at the University of Tokyo
and is now working as a director of Information Processing Group in
RIKEN Frontier Research Program.  He has worked on mathematical
theories of neural networks for thirty years, and his current
interest is, among others, applications of information geometry to
manifolds of neural networks.  He is the past president of the
International Neural Network Society (INNS), a council member of
Bernoulli Society for Mathematical Statistics and Probability, IEEE
Fellow, a member of Scientists Council of Japan, and served as
founding Coeditor-in-Chief of Neural Networks.  He is recipient of
Japan Academy Award, IEEE Emanuel R. Piore Award, IEEE Neural
Networks Pioneer Award, and so on.

----------------------------------------------------------

2) DR. ERIC BAUM:

Dr. Baum remarked that a number of disciplines have independently
reached a near consensus that the brain is a multiagent system
that computes using interaction of modules large compared to
neurons. These different disciplines offer different pictures
of what the modules are and how they interact, and it is
illuminating to compare these different insights. Evolutionary
psychologists talk about modules evolving, as we have evolved
different organs to perform different tasks. They have presented
the Wason selection test, a psychophysical test of reasoning which
seems to indicate that humans have a module specifically for
reasoning about social interactions and cheating detection. Brain
imaging presents a physical picture of modules interacting and will
give great insight into the nature of how modules interact to
compute. Stroke and other lesion victims give insights into
deficits that can arise from damage to a single module. Lakoff and
Johnson's observation that language is metaphorical can be viewed
in modular terms. For example, time is money: you buy time, save
time, invest your time wisely, live on borrowed time, etc. What is
this but a manifestation of a module for valuable resource
management that is applied in different contexts?

Dr Baum also remarked that evolution has built massive knowledge
into us at birth. This knowledge guides our learning. Much of
it is manifested in a detailed intermediate reward function
(pleasure, pain) that guides us to reinforcement learn. There
is copious evidence of built in knowledge-- for example
consider the difference in personalities of a labrador
retriever and a sheepdog. Or for example consider
experiments showing that monkeys, born in the lab without
fear of snakes, can acquire fear of snakes from seeing
a video of a monkey scared of snakes, yet they will not
acquire a fear of flowers from seeing a video of a monkey
recoiling from a flower (Mineka et al, Animal Learning and
Behavior, 8:653 (1980).) Thus learning was a two phase process,
learning during evolution followed by learning during life (and
actually three phase if you consider technology.)

Dr. Baum remarked that traditional neural theories do not
seem to encompass this modular nature well. In his opinion the
critical question in managing interaction of agents is ensuring
that the individual agents all see the correct incentive. This
he feels implies that the multiagent model is essentially an
economic model, and said that he is working in this direction.
Other features of intelligence not well handled by standard
connectionist approaches include metalearning, and metacomputing.
People are able to learn new concepts from a single example, which
requires recursively applying ones knowledge to learning.
Creature's need to be able to decide what to compute and when to
stop computing and act, which again indicates a recursive nature of
intelligence. It is not clear how ever to deal with these problems
in a connectionist framework, but they seem natural within the
context of multiagent economies.

-------------

Eric Baum received B.A. and M.A. degrees in physics from Harvard
University in 1978 and the Ph.D. in physics from Princeton
University in 1982. He has since held positions at Berkeley,
M.I.T., Caltech, J.P.L., and Princeton University and has for
eight years now been a Senior Research Scientist  in the
Computer Science Division of the NEC Research Institute,
Princeton N.J. His primary research interests are in Cognition,
Artificial Intelligence, Computational Learning Theory, and
Neural Networks, but he has also been active in the nascent
field of DNA Based Computers,  co-chairing the first and
chairing the second workshops on DNA Based Computers.
His papers include:
"Zero Cosmological Constant from Minimum Action", Physics Letters
 V 133B, p185 (1983)
"What Size Net Gives Valid Generalization", with D. Haussler,
 Neural Computation v1 (1989) pp148-157
"Neural Net Algorithms that Learn in Polynomial Time from Examples
and Queries", IEEE Transactions in Neural Networks, V2 No. 1 pp
5-19 (1991).
"Best Play for Imperfect Players and Game Tree Search- Part 1
Theory" E. B. Baum and W. D. Smith, (submitted)
"Where Genetic Algorithms Excel", E. B. Baum, D. Boneh, and
C. Garrett, (submitted)
"Toward a Model of Mind as a Laissez-Faire Economy of Idiots,
Extended Abstract", Proceedings of the 13th International
Conference on Machine Learning pp28-36, Morgan Kauffman (1996).

-----------------------------------------------------

3) DR. ROLF ECKMILLER:

Dr. Eckmiller presented three theses about brain-like learning.
First is the notion of factories or modular subsystems. Second, 
the neural networks belong to the geometrical or topological theory
space and not in the algebraic or analytical theory space. Hence
using notions of Von Neumann computing in our analysis might be
equivalent to "barking up the wrong tree." Third, he called upon
the research community to develop a new wave of neural computers -
ones that can adapt weights and time delays, build new layers and
structures, and build and integrate connections between various
parts of the brain. He said that "biological systems are
amathematical" and therefore needs new mathematical tools for
analysis.

-------------

Rolf Eckmiller was born in Berlin, Germany, in 1942. He received
his M.Eng. and Dr. Eng. (with honors) degrees in electrical
engineering from the Technical University of Berlin, in 1967 and
1971, respectively. Between 1967 and 1978, he worked in the fields
of neurophysiology and neural net research at the Free University
of Berlin, and received the habilitation for sensory and
neurophysiology in 1976. From 1972 to 1973 and from 1977 to 1978,
he was a visiting scientist at UC Berkeley and the Smith-Kettlewell
Eye Research Foundation in San Francisco. From 1979 to 1992, he was
professor at the University of Düsseldorf. Since 1992, he has been
professor and head of the Division of Neuroinformatics, Department
of Computer Science at the University of Bonn. His research
interests include vision, eye movements in primates, neural nets
for motor control in intelligent robots, and neurotechnology
with emphasis on retina implants.

----------------------------------------------------------

4) DR. LEE GILES:

Dr. Giles opened his discussion by stating that the connectionist
field has always been a very self-critical one that has always been
receptive to new ideas. Furthermore, the topics proposed here have
been discussed to some extent in the past but are important ones
and certainly worth reevaluation. As an example, the idea of using
memory in learning was one of the earliest ideas in neural networks
and was proposed in the seminal 1943 paper of McCulloch and Pitts.
Today, memory structures are used extensively in neural models
concerned with temporal and sequence anaylsis. For example
recurrent neural networks have successfully been used for such
problems as time series prediction, signal processing, and control.
In discrete time recurrent networks, memory structure and useage
are very important both to the networks performance and
computational power.

Dr. Giles then stated that there is still a great deal of
discrepancy between what our current models can do in theory and
what they can do in practice, and that a deeper foundation for
intelligence processing needs to be established. One approach is to
look at hybrid systems, models that combine many difference
learning and intelligence paradigms - neural networks, AI, etc -
and develop the foundations of intelligent systems by exploring
hybrid system fundamentals. As an example of what intelligent
systems can't do but should be able to, Dr. Giles showed examples
of a pattern classification taken from the book "Pattern
Recognition" by M. Bongard.  Here one sees six examplar
pictures from one class and six from another. The pattern
classification task is to extract the rule(s) that differentiate
one class from the other; a problem that humans can solve but no
machine seems currently seems close to solving. Bongard constructed
100 of these problems. Not only is learning involved but so is
reasoning and explanation. This problem by Bongard is an example of
the types of problems we should be trying to solve and the
questions raised in solving it will give us insights into
constructing and understanding intelligent systems.

Reference: M. Bongard, "Pattern Recognition", Spartan Books, 1970.

----------------

C. Lee Giles is a Senior Research Scienctist in Computer Science at
NEC Research Institute, Princeton, NJ and Adjunct Faculty at the
University of Maryland Institute for Advanced Computer Studies,
College Park, Md. His current research interests are: novel
applications of neural network, machine learning and AI in the WWW,
communications, computing and computers, multi-media, adaptive
control, system identification, language processing, time series
and finance; and dynamically-driven recurrent neural networks -
their computational and processing capabilities and relationships
to other adaptive, learning and intelligent paradigms.  Dr. Giles
was one of the founding members of the Governors Board of the
International Neural Network Society and is a member of the IEEE
Neural Networks Council Technical Committee.  He has served or is
currently serving on the editorial boards of IEEE Transactions on
Neural Networks, IEEE Transactions on Knowledge and Data
Engineering, Journal of Computational Intelligence in Finance,
Journal of Parallel and Distributed Computing, Neural Networks,
Neural Computation, Optical Computing and Processing, Applied
Optics, and Academic Press. Dr. Giles is a Fellow of the IEEE, a
member of AAAI, ACM, INNS, the OSA, and DIMACS - Rutgers University
Center for Discrete Mathematics and Theoretical Computer Science.
Previously, he was a Program Manager at the Air Force Office of
Scientific Research in Washington, D.C.  where he initiated and
managed basic research programs in Neural Networks and in Optics in
Computing and Processing.

-----------------------------------------------------------

5) DR. GEOFREY HINTON:

Dr. Hinton started by pointing out the weaknesses of the
back-propagation algorithm in learning and in certain pattern
recognition tasks. He then focused on the good properties of
Bayesian networks and showed how well the Bayesian networks do on a
certain pattern recognition task.

He believes that prescriptions from well-known researchers about
necessary conditions on biologically realistic learning algorithms
are of some sociological interest but are unlikely to lead to
radically new ideas.

--------

Geoffrey Hinton received his BA in experimental psychology from
Cambridge in 1970 and his PhD in Artificial Intelligence from
Edinburgh in 1978.  He is currently a fellow of the Canadian
Institute for Advanced Research and professor of Computer Science
and Psychology at the University of Toronto.  He does research on
ways of using neural networks for learning, memory, perception and
symbol processing and has over 100 publications in these areas.
He was one of the researchers who introduced the back-propagation
algorithm that is now widely used for practical applications. His
other contributions to neural network research include Boltzmann
machines, distributed representations, time-delay neural nets,
mixtures of experts, and Helmholtz machines. His current main
interest is in unsupervised learning procedures for neural networks
with rich sensory input.  He serves on the editoral boards of
the journals Artificial Intelligence, Neural Computation, and
Cognitive Science.  He is a fellow of the Royal Society of Canada
and of the American Association for Artificial Intelligence and a
former President of the Cognitive Science Society.

---------------------------------------------------------

6) DR. TEUVO KOHONEN:

Dr. Kohonen felt that perhaps we should go back to the basics to
answer some of the questions being raised about connectionist
learning, especially concerning the right forms of transfer
functions and learning laws. He then talked about three levels
of neural functions. At the lowest level, he mentioned the idea
of activation and inhibition as coming from the old views held
in medical science. At the next level, the links between neurons
get modified and change over time. This view was introduced to
the neural science by theorists. He then mentioned that many
earlier and recent neurobiological findings reveal that there
is another third level of control in the brain that controls
the adaptivity of networks, thereby implying certain "nonlocal"
brain mechanisms and their role in designing and training networks.
He called this third level "control of synaptic plasticity" that
was ignored in the past in connectionism. He jokingly mentioned
that his controversial views had developed along different lines
over a long time since he is coming "from another planet" (Finland,
that is). The audience laughed and applauded him heartily.

------------

Teuvo Kohonen, Dr. Eng., Professor of the Academy of Finland, head
of the Neural Networks Research Centre, Helsinki University of
Technolog, Finland. His research areas are associative memories,
neural networks, and pattern recognition, in which he has published
over 200 research papers and four monography books. His fifth book
is on digital computers. Since the 1960s, Professor Kohonen has
introduced several new concepts to neural computing: fundamental
theories of distributed associative memory and optimal associative
mappings, the learning subspace method, the self-organizing feature
maps, the learning vector quantization, and novel algorithms for
symbol processing like the redundant hash addressing and
dynamically expanding context. The best known application of his
work is the neural speech recognition system. Prof. Kohonen has
also done design work for electronics industries. He is recipient
of the Honorary Prize of Emil Aaltonen Foundation in 1983, the
Cultural Prize of the Finnish Commercial Television (MTV) in 1984,
the IEEE Neural Networks Council Pioneer Award in 1991, the
International Neural Network Society Lifetime Achievement Award in
1992, Prize of the Finnish Cultural Foundation in 1994, Technical
Achievement Award of the IEEE Signal Processing Society in 1995,
Centennial Prize of the Finnish Association of Graduate Engineers
in 1996, King-Sun Fu Prize in 1996, and others. He is Honorary
Doctor of the University of York in U.K. and Abo Akademi in
Finland, member of Academia Scientiarum et Artium Europaea, titular
member of the Academie Europeenne des Sciences, des Arts et des
Lettres, member of the Finnish Academy of Sciences and the Finnish
Academy of Engineering Sciences, IEEE Fellow, and Honorary Member
of the Pattern Recognition Society of Finland as well as the
Finnish Society for Medical Physics and Medical Engineering. He was
elected the First Vice President of the International Association
for Pattern Recognition for the period 1982 - 84, and acted as the
first President of the European Neural Network Society during 1991
- 92.

--------------------------------------------------------

7) DR. DANIEL LEVINE:

Dr. Levine agreed that it was indeed time to reconsider the
foundations of connectionist learning. He mentioned that he had
been eager to defend classical connectionist ideas, but then
changed his mind because of some his recent work on analogy
formation and because of work in neuroscience on the role of
neuromodulators and neurotransmitters. He was of the view that
there has to be some "nonlocal" learning mechanisms at work,
particularly because learning of analogies requires that we not
only learn to associate two concepts as in traditional Hebbian
learning, but learn the nature of the association.  (Example:
simply associating Houston to Texas isn't enough to tell us that
Houston is "in" Texas.)  Such nonlocal processes may, he added,
provide more efficient mechanisms for property inheritance and
property transfers.

But Dr. Levine said that reconsidering the foundations of
connectionism does not mean throwing out all existing work but
building on it.  Specifically, connectionist principles such as
associative learning, competition, and resonance, that have been
used in models of pattern recognition and classical conditioning
can also be used in different combinations as building blocks in
connectionist models of more complex cognitive processes.  In these
more complex networks, neuromodulation (via a transmitter
"broadcast" from a distant source node) is likely to play an
important role in selectively amplifying particular subprocesses
based on context signals.

--------------

DANIEL LEVINE is Professor of Psychology at the University of Texas
at Arlington. Dr. Levine holds a Ph.D. in Applied Mathematics from
the Massachusetts Institute of Technology and was a Postdoctoral
Trainee in Physiology at the University of California at Los
Angeles School of Medicine.  His main recent area of research has
been neural network models for the involvement of the frontal lobes
in high-level cognitive tasks and in brain executive function,
including their possible connections with the limbic system and
basal ganglia.  He has also recently published a network model of
the effects of context on preference in multiattribute
decision-making.  Other areas in which he has published include
models of attentional effects in Pavlovian conditioning, dynamics
of nonlinear attractor networks, and models of visual illusions. 
Dr. Levine is author of the textbook, "Introduction to Neural and
Cognitive Modeling," and senior editor of three books that have
arisen out of conferences sponsored by the Dallas-Fort Worth-based
Metroplex Institute for Neural Dynamics (M.I.N.D.).  He has been on
the editorial board of Neural Networks since 1988, serving as Book
Review Editor from 1988 to 1995 and Newsletter editor from 1995 to
the present.  He has been a member of the INNS Board of Governors
since 1995 and current candidate for President-Elect of INNS.  He
is a Program Co-Chair for the International Joint Conference on
Neural Networks in 1997, sponsored by IEEE and INNS.

--------------------------------------------------------

8) DR. JEAN JACQUES SLOTINE:

The issue of learning on an as-needed basis may not have yet
received enough attention. Consider for example a robot
manipulator, initially at rest under gravity forces, and whose
desired task is to just stay there; no control needs being applied
and no adaptation needs occuring, and this is indeed what a good
adaptive controller, whether model-based, parametrized, or "neural"
will do -- actually, doing anything else, e.g. moving so as to
acquire parameter information, would detract it from its task. 
Conversely, if the robot is required to follow a desired trajectory
so complicated that exact trajectory tracking necessarily requires
an exact learning of the robot dynamics, then the guaranteed
tracking convergence of the same adaptive algorithm will
automatically guarantee such learning.  While these issues are now
well understood in a feedback control context, they may be of
interest in a more general setting, since learning seems to
often be equated to learning a whole system model, rather than to a
faster, simpler purely goal-directed learning.

The issue of transmission or computing delays, and the constraints
they impose on stable learning also seem to deserve increased
attention.

------------

Jean-Jacques Slotine was born in Paris in 1959, and received his
Ph.D. from the Massachusetts Institute of Technology in 1983. After
working at Bell Labs in the computer research department, in 1984
he joined the faculty at MIT, where he is now Professor of
Mechanical Engineering and Information Sciences, Professor of Brain
and Cognitive Sciences, and Director of the Nonlinear Systems
Laboratory.  He is the co-author of the textbooks "Robot Analysis
and Control" (Wiley, 1986) and "Applied Nonlinear Control"
(Prentice-Hall, 1991).

-----------------------------------------------------------

9) DR. JOHN TAYLOR:

Dr. Taylor said that having worked at the Brain Institute in
Germany for the last one year, he now has a new and different view
of connectionism. He said that classical connectionism perhaps has
too narrow a view of the brain. He then mentioned that the brain
has a modular structure with three basic regions (nonconcious
regions, concious regions and regions for reasoning,
decision-making and so on). According to him, the following are
some of the important characteristics of the brain:

1) the use of time in discrete chunks, or packets, so that there
are three regimes: one at a few tens of milliseconds, one at the
order of seconds and the third at about a minute.  The first of
these is involved in sensory processing, the second in higher order
processing and the third in frontal 'reasoning'.  The source of
these longer times is as yet unknown but is very important.

2) the effects of neuromodulation from the punishment/reward
system, which provides a global signal,

3) the distribution or break-down of complex tasks into sub-tasks
which are themselves performed by smaller numbers of modules - the
principle of divide and conquer! It is these networks which are now
being uncovered by brain imaging;  how they function in detail will
be the next big task in neuroscience for the next century.

4) the use of a whole battery of neurotransmitters both for ongoing
transmission of information and for learning changes brought about
locally or globally.

He emphasized that memory is indeed used in learning and that in
addition to memory at the higher level (long term memory), there is
working memory and memory in the time delays.

With regard to the issue of global/local learning, he mentioned
that neuromodulation possibly plays a role in passing global
signals. 

As to the question of network design, he said the networks are
designed by the brain and is a function of punishments and rewards
coming from the environment.

In closing, he articulated the view that connectionism should not
be limited to traditional artificial neural networks, but must
include new knowledge being discovered in computational
neuroscience.

--------------

John G. Taylor has been involved in Neural Networks since 1969,
when he developed analysis of synaptic noise in neural
transmission, which has more recently been turned into a neural
chip (the pRAM) with on-chip learning.  He is interested
in a broad range of neural network questions, from theory of
learning and the use of dynamical systems theory and spin glasses
to cognitive understanding  up to consciousness.  He is presently
Director of the Centre for Neural Networks, King's College London 
and a Guest Scientist at the Research Centre Juelich, where he is
involved in developing new tools for analysing brain imaging data
and performing experiments to detect the emergence of
consciousness. He has published over 400 scientific papers in all,
as well as over a dozen books and edited as many again.  He was
INNS President in 1995 and is currently European Editor-in-Chief of
the journal 'Neural Networks', a Governor of INNS and a
Vice-President of the European Neural Network Society.

---------------------------------------------------------

10) DR. DAVID WALTZ:

Dr. Waltz articulated the viewpoint that brains indeed use memory
to learn. He said that we do remember important experiences in life
and then told a story about Marvin Minsky's dog and use of memory
to learn (true story). Minsky's dog had formed the habit of chasing
cars and biting their tires during the regular walks. One day,
while trying to do this, she slipped the leash and got run over and
injured by a car at a certain street corner. From then on, she was
extremely reluctant to go near that particular street corner where
the accident occurred, but continued to chase cars whenever
possible (vivid memories and a wrong learning experience). While
people (and animals) can generally learn better than this, vivid
memories are probably shared by - and important to - all higher
organisms.

Dr. Waltz also emphasized the non-minimal nature of the brain in
the sense that it tries to remember a lot of things in order to
learn. For example, imagine that an intelligent system encounters a
situation that leads to a very negative outcome, and then later
encounters a similar situation that has a positive or neutral
outcome. It is important that enough features of the original
situation be remembered, so that the system can distinguish these
situations in the future, and act accordingly. If the initial
situation is not remembered, but has just been used to make
synaptic weight changes, then the system will have no way to find
features that could distinguish these cases in the future.  So,
with regard to the basic questions, he agreed that we do indeed use
memory to learn in many cases, though not in every case (e.g. motor
skills). On the network design issue, he said that some networks
have been designed through evolution, but that other networks are
indeed designed by the brain through "experience." On global/local
learning, he speculated that perhaps both kinds exist.

---------------

David Waltz is Vice President, Computer Science Research at the NEC
Research Institute in Princeton, NJ, and an Adjunct Professor at
Brandeis University.  From 1984-93, he was Director of Advanced
Information Systems at Thinking Machines Corporation and Professor
of Computer Science at Brandeis. From 1974-83 he was a Professor of
Electical and Computer Engineering at the University of Illinois at
Urbana-Champaign.  Dr. Waltz received SB, SM, and Ph.D. degrees
from MIT, in 1965, 1968, and 1972 respectively.  His research
interests have included constraint propogation, massively parallel
systems for relational and text databases, memory-based reasoning
systems, protein structure prediction using hybrid neural net and
memory-based methods, connectionist models for natural language
processing, and natural language processing. He is President of the
American Association of Artificial Intelligence and was elected a
fellow of AAAI in 1990. He has served as President of ACM SIGART,
Executive Editor of Cognitive Science, AI Editor for Communications
of the ACM.  He is a senior member of IEEE.

---------------------------------------------------------

11) DR. PAUL WERBOS:

Dr. Werbos agreed strongly with the importance of memory-based
learning. He argued that new neural network designs, using
memory-based approaches, could help to solve the classical dilemma
of learning speed versus generalization ability, which has plagued
many practical applications of neural networks. He referred back to
his idea of "syncretism," expressed in chapter 3 of the Handbook of
Intelligent Control and in his paper on supervised learning in
WCNN93 (and Roychowdhury's book). He believes that such mechanisms
are essential to explaining certain capabilities in the neocortex,
reflected in psychology. However, he does not argue that such
mechanisms are present in ALL parts of the brain; for example,
slower learning, based on simpler circuitry, does seem to occur in
motor systems like the cerebellum. Higher motor systems, such as
the neocortex/basal-ganglia/thalamus loops, clearly include
memory-based learning; however, Houk, Ito and others have clearly
shown that some degree of real-time weight-based learning does
exist in the cerebellum. Recent experiments have hinted that even
the cerebellum might be trained in part based on a replay of
memories initially stored in the cerebral cortex; however,
there are reasons to withhold judgment about that idea at the
present time.

Regarding local learning, he stressed that some of the broader
discussions tend to mix up several different notions of "locality,"
each of which needs to be evaluated separately. Massively parallel
distributed processing still remains a fundamental design
principle, both of biological and practical importance. This in no
way rules out the presence of some global signals such as "clocks,"
which are crucial to many designs, and which Llinas and others have
found to be pervasive in the brain.  Likewise, it does not rule
out subsystems for "growing" and "pruning" connections, which are
already well established in the connectionist literature
(discussed in chapter 10 of the Handbook of Intelligent Control,
and in many other places.).

Regarding the role of learning versus evolution, he does not see
the same kind of "either-or" choice that many people assume. His
views are expressed in detail in a new paper to appear in Pribram's
new book on Values from Erlbaum, 1997, reflecting the new "decision
block" or "three brain" design discussed at this conference.

--------

Dr. Paul J. Werbos holds 4 degrees from Harvard University and the
London School of Economics, covering economics, mathematical
phyiscs, decision and control, and the backpropagation algorithm.
His 1974 PhD thesis presented the true backpropagation algorithm
for the first time, permitting the efficient calculation of
derivatives and adaptation of all kinds of nonlinear sparse
structures, including neural networks; it has been reprinted in its
entirety in his book, The Roots of Backpropagation, Wiley,
1994, along with several related seminal and tutorial papers. In
these and other more recent papers, he has described how 
backpropagation may be incorporated into new intelligent control
designs with extensive parallels to  the structure of the human
brain. See the hot links on www.nsf.gov/eng/ecs/enginsys.htm

      Dr. Werbos runs the Neuroengineering program and the SBIR
Next Generation Vehicle program at the National Science Foundation.
He is Past President of the International Neural Network
Society(INNS), and is currently on the governing boards both of
INNS and of the IEEE society for Systems, Man and Cybernetics..
Prior to NSF, he worked at the University of Maryland and the U.S.
Department of Energy. He was born in 1947 near Philadelphia,
Pennsylvania, has three children, and attends Quaker meetings. His
publications range from neural networks through to quantum
foundations, energy economics, and issues of consciousness.
---------------------------------------------------