There is a continuing discussion on the above topic over several
mailing lists. I thought neur-sci members might also want to
participate in this discussion. So please post. I will compile all
responses and post it to the participating mailing lists.
================================================================
I thought it might be productive to the discussion if I compiled
and posted the responses I have received so far. I am posting them
without
any comments. I hope this will promote further discussion on this
topic. The original posting is attached below for reference.
============================================================
Response # 1:
Re: your note on plasticity and learning
Asim,
As you mention, neuroscience tends to equate network plasticity
with learning. Connectionists tend to do the same. However this
raises a problem with biological systems because this conflates the
processes of
development and learning. Even the smartest organism starts from an
egg, and develops for its entire lifespan - how do we distinguish
which
changes are learnt, and which are due to development. No one would
argue that we
*learn* to have a cortex, for instance, even though it is due to
massive
emryological changes in the central nervous system of the animal.
This isn't a problem with artificial nets, because they do not
usually have a true developmental process and so there can be no
confusion between the two; but it has been a long-standing problem
in the ethology
literature, where learnt changes are contrasted with "innate"
developmental
ones. A very interesting recent contribution to this debate is
Andre
Ariew's "Innateness and Canalization", in Philosophy of Science 63
(Proceedings), in which he identifies non-learnt changes as being
due to canalised
processes. Canalization was a concept developed by the biologist
Waddington in
the 40's to describe how many changes seem to have fixed end-goals
that are robust against changes in the environment.
The relationship between development and learning was also
thoroughly explored by Vygotsky (see collected works vol 1, pages
194-210).
I'd like to see what other sorts of responses you get,
Joe Faith <josephf at cogs.susx.ac.uk>
Evolutionary and Adaptive Systems Group,
School of Cognitive and Computing Sciences,
University of Sussex, UK.
=================================================================
Response # 2:
I fully agree with you, that local learning is not the one
and only ultimate approach - even though it results in very
good learning for some domains.
I am currently writing a paper on the competitive learning
paradigm. I am proposing, that this competition that occurs
e.g. within neurons should be called local competition. The
network as a whole gives a global common goal to these local
competitors and thus their competition must be regarded as
cooperation from a more global point of view.
There is a nice paper by Kenton Lynne that integrates the
ideas of reinforcement and competition. When external
evaluations are present, they can serve as teaching values,
if nor the neurons compete locally.
@InProceedings{Lynne88,
author = {K.J.\ Lynne},
title = {Competitive Reinforcement Learning},
booktitle = {Proceedings of the 5th International Conference
on
Machine Learning},
year = {1988},
publisher = {Morgan Kaufmann},
pages = {188--199}
}
Best regards,
Christoph Herrmann
----------------------------------------------------------
Christoph Herrmann Visiting researcher
Hokkaido University
Meme Media Laboratory
Kita 13 Nishi 8, Kita- Tel: +81 - 11 - 706 - 7253
Sapporo 060 Fax: +81 - 11 - 706 - 7808
Japan Email: chris at meme.hokudai.ac.jphttp://aida.intellektik.informatik.th-darmstadt.de/~chris/
----------------------------------------------------------
=============================================================
Response #3:
I've just read your list of questions on local vs. global learning
mechanisms. I think I'm sympathatic to the implications or
presuppositions of your questions but need to read them more
carefully later. Meanwhile, you might find very interesting a
two-part article on such a mechanism by Peter G. Burton in the 1990
volume
of _Psychobiology_ 18(2).119-161 & 162-194.
Steve Chandler
<chandler at uidaho.edu>
===============================================================
Response #4:
A few years back, I wrote a review article on issues
of local versus global learning w.r.t. synaptic plasticity.
(Unfortunately, it has been "in press" for nearly 4 years).
Below is an abstract. I can email the paper to you in TeX or
postscript format, or mail you a copy, if you're interested.
Russell Anderson
------------------------------------------------
"Biased Random-Walk Learning:
A Neurobiological Correlate to Trial-and-Error"
(In press: Progress in Neural Networks)
Russell W. Anderson
Smith-Kettlewell Eye Research Institute
2232 Webster Street
San Francisco, CA 94115
Office: (415) 561-1715
FAX: (415) 561-1610
anderson at skivs.ski.org
Abstract:
Neural network models offer a theoretical testbed for
the study of learning at the cellular level.
The only experimentally verified learning rule,
Hebb's rule, is extremely limited in its ability
to train networks to perform complex tasks.
An identified cellular mechanism responsible for
Hebbian-type long-term potentiation, the NMDA receptor,
is highly versatile. Its function and efficacy are
modulated by a wide variety of compounds and conditions
and are likely to be directed by non-local phenomena.
Furthermore, it has been demonstrated that NMDA receptors
are not essential for some types of learning.
We have shown that another neural network learning
rule, the chemotaxis algorithm, is theoretically much more powerful
than Hebb's rule and is consistent with experimental data.
A biased random-walk in synaptic weight space is
a learning rule immanent in nervous activity and
may account for some types of learning -- notably the
acquisition of skilled movement.
------------------------------------------
E-mail:
send request to rwa at milo.berkeley.edu
Slow mail:
Russell Anderson
2415 College Ave. #33
Berkeley, CA 94704
==========================================================
Response #5:
Asim Roy typed ...
>> B) "Pure" local learning does not explain a number of other
> activities that are part of the process of learning!!
....
>> So, in the whole, there are a "number of activities" that need to
> be
> performed before any kind of "local learning" can take place.
These
> aforementioned learning activities "cannot" be performed by a
> collection of "local learning" cells! There is more to the
process
> of learning than simple local learning by individual cells. Many
> learning "decisions/tasks" must precede actual training by "local
> learners." A group of independent "local learners" simply cannot
> start learning and be able to reproduce the learning
> characteristics and processes of an "autonomous system" like the
> brain.
I cannot see how you can prove the above statement (particularly
the last sentence). Do you have any proof. By analogy, consider
many insect colonies (bees, ants etc). No-one could claim that one
of
the insects has a global view of what should happen in the colony.
Each
insect has its own purpose and goes about that purpose without
knowing the
global purpose of the colony. Yet an ants nest does get built, and
the colony does survive. Similarly, it is difficult to claim that
evolution has a master plan, order just seems to develop out of
chaos.
I am not claiming that one type of learning (local or global) is
better than another, but I would like to see some evidence for your
somewhat outrageous claims.
> Note that the global learning mechanism may actually be
implemented
> with a collection of local learners!!
You seem to contradict yourself here. You first say that local
learning cannot cope with many problems of learning, yet global
learning can. You then say that global learning can be implemented
using
local learners. This is like saying that you can implement things
in C,
that cannot be implemented in assembly!! It may be more convenient
to implement
it in C (or using global learning), but that doesn't make it
impossible for assembly.
Cheers,
Brendan.
-------------------------------------------------------------------
Brendan McCane, PhD. Email:
mccane at cs.otago.ac.nz
Comp.Sci. Dept., Otago University, Phone: +64 3 479 8588.
Box 56, Dunedin, New Zealand. There's only one catch -
Catch 22.
===============================================================
Response #6:
In regards to arguments against global learning:
I think no one seriously questions this possibility,
but think that global learning theories are currently
non-verifiable/ non-falsifyable.
Part of the point of my paper was that there ARE ways to
investigate non-local learning, but it requires changes
in current experimental protocols.
Anyway, good luck.
I look forward to seeing your compilation.
Russell
------------------------------------------
E-mail:
send request to rwa at milo.berkeley.edu
Slow mail:
Russell Anderson
2415 College Ave. #33
Berkeley, CA 94704
==============================================================
Response #7:
I am sorry that it has taken so long for me to reply to
your inquiry about plasticity and local/global learning. As I
mentioned in my first note to you, I am sympathetic to the view
that learning
involves some sort of overarching, global mechanism even though the
actual information storage may consist of distributed patterns of
local
information. Because I am sympathetic to such a view, it makes it
very difficult for me to try to imagine and anticipate the problems
for such views. That's why I am glad to see that you are
explicitly
trying to find people to point out possible problems; we need the
reality
check.
The Peter Burton articles that I have sent you describes
exactly the kind of mechanism implied by your first question: Does
plasticity imply local learning? Burton describes a neurological
mechanism by which local learning could emerge from a global
signal.
Essentially he posits that whenever the new perceptual input being
attended to at
any given moment differs sufficiently from the record of previously
recorded experiences to which that new input is being compared, the
difference triggers a global "proceed-to-store" signal. This
signal creates a neural "snapshot" (my term, not Burton's) of the
cortical
activations at that moment, a global episodic memory (subject to
stimulus sampling
effects, etc.). Burton goes on to describe how discrete episodic
memories could become associated with one another so as to give
rise to
schematic representations of percepts (personally I don't think
that positing
this abstraction step is necessary, but Burton does it).
As neuroscientists sometimes note, while it is widely
assumed that LTP/LTD are local learning mechanisms, the direct
evidence for such a hypothesis is pretty slim at best. Of course
of of the most
serious problems with that view is that the changes don't last very
long and thus are not really good candidates for long term (i.e.,
life
long) memory. Now, to my mind, one of the most important
possibilities
overlooked in LTP studies (inherently so in all in vitro
preparations and so
far as I know--which is not very far because this is not my
field--in the in
vivo preparations that I have read about) is that LTP/D is either
an artifact of the experiment or some sort of short term change
which
requires a global signal to become consolidated into a long term
record.
Burton describes one such possible mechanism.
Another motivation for some sort of global mechanism comes
from the so-called 'binding problem' addressed especially by the
Damasio's, but others too. Somehow somewhere all the distributed
pieces of information about what an orange is, for example, have to
be tied
together. A number of studies of different sorts have demonstarted
repeatedly
that such information is distributed throughout cortical areas.
Burton distinguishes between "perceptual learning"
requiring no external teacher (either locally or globally) and
"conceptual learning", which may require the assistance of a
'teacher'. In his
model though, both types of learning are activated by global
"proceed-to-learn" signals triggered in turn by the global
summation of local
disparities between remembered episodes and current input.
I'll just mention in closing that I am particularly
interested in the empirical adequacy of neuropsychological accounts
such as Burton's because I am very interested in "instance-based"
or
"exemplar-based" models of learning. In particular, Royal
Skousen's _Analogical Modeling of Language_ (Kluwer, 1989)
describes an explicit,
mathematical model for predicting new behavior on analogy to
instances stored in
long term memory. Burton's model suggests a possible neurological
basis for
such behavior.
==============================================================
Response #8:
*******************************************************************
Fred Wolf E-Mail:
fred at chaos.uni-frankfurt.de
Institut fuer Theor. Physik
Robert-Mayer-Str. 8 Tel: 069/798-23674
D-60 054 Frankfurt/Main 11 Fax: (49) 69/798-28354
Germany
*******************************************************************
Dear Asim Roy,
could you please point me to a few neuroBIOLOGICAL references
that justify your claim that
>> A predominant belief in neuroscience is that synaptic plasticity
> and LTP/LTD imply local learning (in your sens).
>
I think many people appreciate that real learning implies the
concerted interplay of a lot of different brain systems and should
not even be attempted to be explained by "isolated local learners".
See e.g. the series of review-papers on memory in a recent volume
of PNAS 93 (1996) (http://www.pnas.org/).
Good luck with your general theory of global/local learning.
best wishes
Fred Wolf
==============================================================
Response #9:
Comp-Neuro Mailing List wrote:
>>===================================================================
> Date: Sun, 16 Feb 1997 22:57:11 -0500 (EST)
> From: Asim Roy <ataxr at IMAP1.ASU.EDU>
> Subject: Does plasticity imply local learning? And other
questions
>> Any comments on these ideas and possibilities are welcome.
>> Asim Roy
> Arizona State University
I am into neurocomputing for several years. I read your arguments
with interest. They certainly deserve further attention. Perhaps
some combination of global-local learning agents would be the right
choice.
- Vassilis G. Kaburlasos
Aristotle University of Thessaloniki, Greece
==============================================================
===============================================================
Original Memo:
A predominant belief in neuroscience is that synaptic plasticity
and LTP/LTD imply local learning. It is a possibility, but it is
not the only possibility. Here are some thoughts on some of the
other possibilities (e.g. global learning mechanisms or a
combination of global/local mechanisms) and some discussion on the
problems associated with "pure" local learning.
The local learning idea is a very core idea that drives research in
a number of different fields. I welcome comments on the questions
and issues raised here.
This note is being sent to many listserves. I will collect all of
the responses from different sources and redistribute them to all
of the participating listserves. The last such discussion was very
productive. It has led to the realization by some key researchers
in the connectionist area that "memoryless" learning perhaps is not
a very "valid" idea. That recognition by itself will lead to more
robust and reliable learning algorithms in the future. Perhaps a
more active debate on the local learning issue will help us resolve
this issue too.
A) Does plasticity imply local learning?
The physical changes that are observed in synapses/cells in
experimental neuroscience when some kind of external stimuli is
applied to the cells may not result at all from any specific
"learning" at the cells. The cells might simply be responding to a
"signal to change" - that is, to change by a specific amount in a
specific direction. In animal brains, it is possible that the
"actual" learning occurs in some other part(s) of the brain, say
perhaps by a global learning mechanism. This global mechanism can
then send "change signals" to the various cells it is using to
learn a specific task. So it is possible that in these neuroscience
experiments, the external stimuli generates signals for change
similar to those of a global learning agent in the brain and that
the changes are not due to "learning" at the cells themselves.
Please note that scientific facts and phenomenon like LTP/LTD or
synaptic plasticity can probably be explained equally well by many
theories of learning (e.g. local learning vs. global learning,
etc.). However, the correctness of an explanation would have to be
judged from its consistency with other behavioral and biological
facts, not just "one single" biological phenomemon or fact.
B) "Pure" local learning does not explain a number of other
"activities" that are part of the process of learning!!
When learning is to take place by means of "local learning" in a
network of cells, the network has to be designed prior to its
training. Setting up the net before "local" learning can proceed
implies that an external mechanism is involved in this part of the
learning process. This "design" part of learning precedes actual
training or learning by a collection of "local learners" whose only
knowledge about anything is limited to the local learning law to
use! In addition, these "local learners" may have to be told what
type of local learning law to use, given that a variety of
different types can be used under different circumstances. Imagine
who is to "instruct and set up" such local learners which type of
learning law to use? In addition to these, the "passing" of
appropriate information to the appropriate set of cells also has
to be "coordinated" by some external or global learning mechanism.
This
coordination cannot just happen by itself, like magic. It has to be
directed from some place by some agent or mechanism.
In order to learn properly and quickly, humans generally collect
and store relevant information in their brains and then "think"
about it (e.g. what problem features are relevant, complexity of
the problem, etc.). So prior to any "local learning," there must be
processes in the brain that "examine" this "body of
information/facts" about a problem in order to design the
appropriate network that would fit the problem complexity, select
the problem features that are meaningful, etc. It would be very
difficult to answer the questions "What size net?" and "What
features to use?" without looking at the problem (body of
information)in great detail. A bunch of "pure" local learners,
armed with their local learning laws, would have no clue to these
issues of net design, generalization and feature selection.
So, in the whole, there are a "number of activities" that need to
be performed before any kind of "local learning" can take place.
These aforementioned learning activities "cannot" be performed by a
collection of "local learning" cells! There is more to the process
of learning than simple local learning by individual cells. Many
learning "decisions/tasks" must precede actual training by "local
learners." A group of independent "local learners" simply cannot
start learning and be able to reproduce the learning
characteristics and processes of an "autonomous system" like the
brain.
Local learning or local computation, however, is still a feasible
idea, but only within a general global learning context. A global
learning mechanism would be the one that "guides" and "exploits"
these local learners or computational elements. However, it is also
possible that the global mechanism actually does all of the
computations (learning) and "simply sends signals" to the network
of cells for appropriate
synaptic adjustment. Both of these possibilities seem logical: (a)
a "pure" global mechanism that learns by itself and then sends
signals to the cells to adjust, or (b) a global/local combination
where the global mechanism performs certain tasks and then uses the
local mechanism for training/learning.
Thus note that the global learning mechanism may actually be
implemented with a collection of local learners or computational
elements!! However, certain "learning decisions" are made in the
global sense and not by "pure" local learners.
The basic argument being made here is that there are many tasks in
a "learning process" and that a set of "local learners" armed with
their local learning laws is incapable of performing all of those
tasks. So local learning can only exist in the context of global
learning and thus is only "a part" of the total learning process.
It will be much easier to develop a consistent learning theory
using the global/local idea. The global/local idea perhaps will
also give us a better handle on the processes that we call
"developmental" and "evolutionary." And it will, perhaps, allow us
to better explain many of the puzzles and inconsistencies in our
current body of discoveries about the brain. And, not the least, it
will help us construct far better algorithms by removing the
"unwarranted restrictions" imposed on us by the current ideas. Any
comments on these ideas and possibilities are welcome.
Asim Roy
Arizona State University
