Copy Link
Add to Bookmark
Report
Neuron Digest Volume 07 Number 41
Neuron Digest Friday, 19 Jul 1991 Volume 7 : Issue 41
Today's Topics:
European Neural Network Society
EUROPEAN NEURAL NETWORK SOCIETY
Hybrid systems -- what commercial packages exist?
BIOPROP simulator announced
Re: First Impressions of Bions
Re: Neural Chess - Board Setup
request for reader input
Introduction and request for infromation
Want info on "AI on Wall Street conference"
Stability proofs for recurrent networks
Send submissions, questions, address maintenance and requests for old issues to
"neuron-request@hplabs.hp.com" or "{any backbone,uunet}!hplabs!neuron-request"
Use "ftp" to get old issues from hplpm.hpl.hp.com (15.255.176.205).
----------------------------------------------------------------------
Subject: European Neural Network Society
From: R09614%BBRBFU01.BITNET@CUNYVM.CUNY.EDU
Date: Wed, 03 Jul 91 17:19:59 +0200
PRESS RELEASE
During ICANN91, the International Conference on Artificial Neural
Networks, held in Helsinki, June 24-28, 1991, ENNS, the European Neural
Network Society has been created.
Its main objectives are to sponsor furture ICANN conferences and organize
summer schools in the area, and to develop a quarterly newsletter and
electronic bulletin board that will keep members informed of developments
in the field of artificial and biological neural networks. It is
expected that Special Interest Groups will be formed in the future under
the umbrella of the society.
The new society welcomes membership worldwide from those interested in
Neural Network research. One of the benefits will be de reduction of
fees for all future activities of the society.
The current officiers of the ENNS are:
President Tuevo Kohonen
Helsinki University of Technology
Espoo, Finland
Vice Presidents Igor Aleksander
Imperial College London
London, U.K.
Rolf Eckmiller
Heinrich Heine University
Dusseldorf, F.R.G.
John Taylor
King's College London
London, U.K.
Secretary Agnessa Babloyantz
Universite Libre de Bruxelles
Brussels, Belgium
Treasurer Rodney Cotterill
Technical University of Denmark
Lyngby, Denmark
For further information, please contact A. Babloyantz at:
Universite Libre de Bruxelles Phone: 32 - 2 - 650 55 40
CP 231 - Campus Plaine, Fax: 32 - 2 - 650 57 67
Boulevard du Triomphe, E-mail: adestex @ bbrbfu60.bitnet
B-1050 Bruxelles,
BELGIUM
------------------------------
Subject: EUROPEAN NEURAL NETWORK SOCIETY
From: R09614%BBRBFU01.BITNET@CUNYVM.CUNY.EDU
Date: Fri, 05 Jul 91 18:57:43 +0200
EUROPEAN NEURAL NETWORK SOCIETY
Due to connection problems, some E-mails did not arrive to the address
given in the first announcement. Please send requests or comments for
ENNS at the following address:
R09614@BBRBFU01.bitnet
If you had already sent a message, please re-send a copy to this address.
Some files have been lost.
Thank you
------------------------------
Subject: Hybrid systems -- what commercial packages exist?
From: "CHAOCHANG WITH NEW TEL. 301-2474657 & NEW ADDR. #20 CASEY COURT , BALTIMORE, MD 21228" <chiu@umbc2.umbc.edu>
Date: 08 Jul 91 21:54:00 -0400
Dear Sir:
I would like to ask the help from you about the current commercial
software pakages that include both expert rule based processing and
neural network capabilities. Or the neural network system can be embedded
in expert system vice vesa. You assistance will be greatly appreciated.
Best Regards
Chaochang Chiu
------------------------------
Subject: BIOPROP simulator announced
From: smuskal%calv01.hepnet@Csa1.LBL.Gov (Steven Marc Muskal)
Date: Tue, 09 Jul 91 11:44:31 -0700
We have written a generalized neural network simulator, BIOPROP,
which can automate much of feed-forward network training and testing for
2 and 3 layer networks. Within the interactive simulator, the user can
define his own variables, loop and condition on these variables, etc.
BIOPROP uses either traditional backpropagation or conjugate-gradient
backpropagation for training and can be tested as a continuous output
device, "winner-take- all" classifier, or "multiple category" classifier.
The software is written in FORTRAN for a SGI 4D system (UNIX), but has
been implemented on other UNIX systems (Stardent, Convex, etc). We have
written a comprehensive manual, complete with example scripts that can be
made available in Microsoft Word (mac/IBM versions) and/or postscript
form.
If you are interested in obtaining this software, send your name,
institution, phone, e-mail, and regular mail address to either:
Steven Muskal
Laboratory of Biodynamics
University of California at Berkeley
Berkeley, CA 94709
(415)-486-4338
or
smuskal@sb1.cchem.berkeley.edu
or
smuskal@csa1.lbl.gov
------------------------------
Subject: Re: First Impressions of Bions
From: ring@cs.utexas.edu (Mark Ring)
Date: Thu, 11 Jul 91 15:48:42 -0500
At the beginning of June, Stephen Smoliar raised some interesting
comments regarding a paper I announced, ftp'able from ohio-state
(ring.ml91.ps.Z). Having been out of town for 4 of the past 5 weeks, I'd
like now to respond to some of these concerns.
> It was with a fair amount of curiosity that I responded to Mark
> Ring's announcement of the availability of his paper, "Incremental
> Development of Complex Behaviors through Automatic Construction of
> Sensory-motor Hierarchies." Papers which come on that strong often
> tend to deliver far less than they promise. This one seems to have
> hooked my curiosity, however; and, in the interest of trying to
> drum up some discussion, I wanted to raise some questions basically
> inspired by the final sections on related work and conclusions.
> In the case of citing related work, I think that Ring may have
> committed two "sins of omission" (which are probably not even major
> enough to be called "sins"). The more important is that the
> approach he advocates seems quite consisted with the approach which
> Minsky seems to be trying to pursue in THE SOCIETY OF MIND. I
> think there is a good chance that bions are the sorts of
> computational elements from which Minsky-like agents might
> ultimately be built; and I, for one, would be interested in some
> discussion on this observations.
I've been reading The Society of Mind in order to give a more educated
response here, but I'm not finished yet. I see a lot of similarities and
possible relationships, but my initial assessment is that bions are
probably not the devices Minsky requires. They do represent behaviors as
hierarchies as his agents do, and they're formed in ways mildly
resembling the way K-line hierarchies might be formed (sections 8.8 -
8.10); they allow the incremental development he refers to (section
8.11); and there is a resemblance later on (27.2 and 27.3) to the gating
function of "suppressors" and "sensors".
Obvious differences include the specifics of K-lines (chapter 8), safely
learning hierarchies (10.9), the A-brain, B-brain dichotomy (6.4) and
many other issues that are not addressed or are not needed by bions.
More generally, it seems from Minsky's remarks that there is probably no
single framework underlying the agents he has in mind. ("What magical
trick makes us intelligent? *The trick is that there is no trick.* The
power of intelligence stems from our vast diversity, not from any single,
perfect principle." (section 30.8) ) He emphasizes that the agencies are
heterogeneous and specialized. A system of bions, however, does have a
single set of principles that are applied uniformly. On the other hand,
most of the pictures Minsky draws resemble connectionist networks, so
maybe there is some underlying description he might say all agencies
could share. But I would be surprised (though very interested) if it
bore more than a superficial resemblance to a bion.
> Another point involving related work is that I suspect the
> resemblance to LISP (and there is no denying that resemblance) is
> more than superficial. Basically, LISP taught us that an ordered
> pair of pointers is all you need to build hierarchical structures
> of arbitrary complexity. Therefore, it should come as no surprise
> that the sorts of hierarchies which Ring is interested in
> developing should end up being built of the same basic stuff with
> only minor variations applied to allow for a different regimen of
> control.
It is true that bions make use of a binary-tree structure in order to
build arbitrarily complex hierarchies. LISP is somewhat more
sophisticated, though, since by allowing recursion it actually uses
directed (possibly cyclic) graphs instead of just trees (directed acyclic
graphs). But there are also some other, not-so-minor differences. Take
for example the real-valued connection weights, the gated connections,
and the hidden connections to a bion from its condition and consequent
(its two children). There is also a specific temporal semantics imposed
on the bion's two children. Now, it could be that these differences are
mostly necessitated by the control regimen, but don't forget how drastic
the differences in control are between LISP and recurrent connectionist
networks. Like the units in connectionist networks, every bion operates
under identical principles, and is updated at every time-step. The
control of the hierarchies, therefore, might best be thought of as a
property that simply emerges from this updating process that's
continually applied to the network.
> My greatest disappointment was that the one example which was
> invoked in order to demonstrate bions was just not very realistic.
> I would be very interested in just how much has been done with
> bions so far. Can they, for example, be invoked for the sort of
> learning one might wish to associate with a Brooks- like robot
I'll have to agree that the example is not very realistic. This is
because, first: though it is representative of those hierarchies the
system has constructed so far, it was intended less as a description of
what bions are designed for than as a clarification for the reader of a
simple bion hierarchy's operation. Second, (as was stated in the paper)
the system is currently being developed in an attempt to address four
specific needs in the connectionist and machine-learning communities: 1)
incremental development, 2) history limitation, 3) behavior organization,
and 4) credit assignment among behaviors. Hopefully, though the initial
hierarchies I'm using for testing are tiny, they capture the important
characteristics needed in larger, more complex and interesting behaviors.
Perhaps, for example, bions will be able to construct behavior
hierarchies reminiscent of those Brooks uses in his robots. This has
been suggested to me before, and it's a question I'd like to answer; but
for now, it's something that requires more research.
> Another question concerns this whole premise of incremental
> development. Being able to build hierarchies from behaviors which
> are already in your repertoire is certainly a good thing, but where
> do you start? Where does an agent's initial set of behaviors come
> from? My inclination would be to turn to the work of Edelman for
> an answer here and consider the possibility of a selectionist
> mechanism working of shaking down an extremely vast repertoire of
> behaviors, sifting it down to a viable set of building blocks.
The behaviors begin at whatever level the implementor might wish them to
begin. If the implementor wishes to develop sophisticated behaviors by
hand (either by programming them explicitly or by building hand-
constructed bion hierarchies), these might form the initial building
blocks for learning. But complicated behaviors at the lowest-level
aren't required. Simple behaviors, which merely send signals to robot
actuators or receive them from robot sensors, should probably be
sufficient. Again, the idea is: it shouldn't matter where learning
begins, provided the system can *learn* (perhaps incrementally) all the
requisite component behaviors it needs in order then to learn the desired
task.
> I am also a bit concerned about putting too much "chunking" effort
> into a single bion. A "chunked" bion can, indeed, encapsulate a
> rather complex pattern of behavior; but that behavior may involve
> some number of low-level decisions which have to be satisfied
> before it can terminate. (This is even apparent in the relatively
> simple wall-avoiding example which is given.) The point is that
> what Ring chooses to call "intention" may be based on a set of
> assumed conditions, and the discovery that some of those conditions
> do not hold may only arise while trying to satisfy the intention.
> In other words there may be grounds for a "chunked" bion to start
> up; but those grounds to not guarantee that it will ever terminate.
> I think it is necessary to consider to what extent this is an
> important issue and would welcome both a response from Ring and
> further discussion on the matter.
This is an important point not addressed in the paper: what happens if
expected components of a chosen behavior fail? These situations should,
in fact, be handled properly by the bion system. The behavior stops
because the conditions necessary for it to continue are missing. At that
point the agent (system of bions) is in exactly the same situation it
would have been in when it was learning lower-level behaviors. This is
as it should be: if there is a failure of expectancy at some level in the
behavior, the agent should learn how to deal with the novel situation at
that level. So, there are two aspects here: 1) the process of behavior
execution should continue from the lower level where the higher-level
behavior selection went wrong, and 2) learning should also continue at
this point to allow the construction of modified behaviors that take the
unexpected conditions into account. The latter is essential if the agent
is to continue learning about its environment, or if behavior hierarchies
become damaged through weight changes resulting from disuse.
Mark Ring
ring@cs.utexas.edu
------------------------------
Subject: Re: Neural Chess - Board Setup
From: David Kanecki <kanecki@vacs.uwp.edu>
Date: Sat, 13 Jul 91 09:17:48 -0500
Re: Neural Chess - Board Set up
Basis: Confusion from responses of following game
In my chess notation system the board setup was as follows:
A B C D E F G H
=================================
1 |WR |WN |WB |WK |WQ |WB |WN |WR |
|-------------------------------|
2 |WP |WP |WP |WP |WP |WP |WP |WP |
|-------------------------------|
3 | | | | | | | | |
---------------------------------
4 | | | | | | | | |
---------------------------------
5 | | | | | | | | |
---------------------------------
6 | | | | | | | | |
---------------------------------
7 |BP |BP |BP |BP |BP |BP |BP |BP |
---------------------------------
8 |BR |BN |BB |BQ |BK |BB |BN |BR |
=================================
Thus, the black king is at E8 and the black queen is at D8 resulting in
a unconventional setting that would test the ability of the program to
think its way to a solution.
I claim to have developed a thinking intelligent chess program using neural
networks. This program is over 100 pages long and has been tested many
times with conventional and unconvential setups which I will report on a
time to time basis.
David H. Kanecki, Bio. Sci., A.C.S.
kanecki@vacs.uwp.wisc.edu
------------------------------
Subject: request for reader input
From: mwitten@hermes.chpc.utexas.edu
Date: Sun, 14 Jul 91 18:07:16 -0500
Dear Reader:
I am writing this note to solicit reprints/preprints for a book that I am
writing. The book is called Frankenstein In The Machine (Pergamon Press).
It is aimed at being a comprehensive study of the role of high
performance computing in medicine/biology/dentistry/allied health
sciences. The book is in its final stages of organization. I am trying to
make the book relatively comprehensive. I am looking for the following
items:
(1)Color pictures/black and white graphics (slides/prints) of actual
results. Same areas are
-medical imaging
-molecular visualization
-simulation visualization
-microscopy
-medical visualization
but are not limited to these areas. I will make copies of anything
sent to me and gladly return it to you. In addition, I will
acknowledge all figures used in the book. If you send pictures, please
include appropriate citation/legend
(2)Copies of reprints/preprints in the area of computational medicine,
dentistry, veterinary medicine, and allied health sciences. Areas
include, not exclusively,
-computational physiology and metabolism
-computational pharmacology, pharmacodynamics, drug design
-computational chemistry
-demography, epidemiology, and statistics/biostatistics
-disease modeling
-cell biology
-radiology
-surgery
-cardiology
-renal function, brain modeling, neurophysiology, bone modeling
joint modeling
-issues in teaching
(3)Names and addresses of individuals working in areas related to the
above and who might be interested in contributing to it.
I am trying to make this as complete a book as possible and do not wish
to commit the faux pas of forgetting someone, if at all possible. Please
feel free to circulate copies of this note, to repost copies of this
note, and to discuss the book with others. I am setting an end of July
1991 target for receipt of materials. However, I am willing to be
flexible if necessary. If you are interested in being kept informed about
this project, please also let me know.
My contact information is below. Please feel free to contact me if you
have any further questions.
Matthew Witten, Ph.D.
Director, Applications Research & Development
Associate Director,
UT System Center For High Performance Computing
Balcones Research Center, 1.154 CMS
10100 Burnet Road, Austin, TX 78758-4497 USA
Phone: (512) 471-2472 FAX: (512) 471-2445/2449
E-MAIL MWITTEN@HERMES.CHPC.UTEXAS.EDU
or
MWITTEN@UTCHPC.BITNET
_____________________________________________________________________
Matthew Witten, Ph.D.
Director, Applications Research & Development
Associate Director,
UT System Center For High Performance Computing
Balcones Research Center, 1.154 CMS
10100 Burnet Road, Austin, TX 78758-4497 USA
Phone: (512) 471-2472 FAX: (512) 471-2445/2449
E-MAIL MWITTEN@CHPC.UTEXAS.EDU
or
MWITTEN@UTHERMES.BITNET
"some intellectuals make their living by creating
obscurities for the rest of us to puzzle over"
_____________________________________________________________________
------------------------------
Subject: Introduction and request for infromation
From: Norian%fee.unicamp.ansp.br@UICVM.uic.edu
Date: Tue, 16 Jul 91 10:34:11 -0500
Dear Sir;
We are sending below an abstract of the work we are currently
developing. We would like to receive information about the following
aspects of artificial neural networks:
- software simulations;
- hardware implementation (analog/digital & electronic/optical);
- algorithms;
- content addressable memories;
- threshold logic;
- holography;
- parallel distributed processing.
Abstract
========
Artificial Neural Netwoks (ANNs) are characterized by massively parallel
processing and very high connectivity among the processing cells. This
implies very intensive inter-elements communication and learnig
algorithms to determine their dynamic behavior. In this work we are
studying several models of ANNs used as content addressable memories.
Presently we are simulating the Hopfield model on a Transputer-based
machine configured as a hypercube. The outcomes of these simulations are
expected to give us a better knowledge of the stability conditions and
storage capacity of the net.
Sincerely yours.
Marcelo Jara and Norian Marranghello.
e_mail: norian@fee.unicamp.ansp.br
------------------------------
Subject: Want info on "AI on Wall Street conference"
From: MSANDRI%IVRUNIV.BITNET@ICINECA.CINECA.IT
Date: Tue, 16 Jul 91 19:02:43 -0100
I have heard about the First International Conference on Artificial
Intelligence Applications on Wall Street (1991), whose proceedings are
published by the IEEE Computer Society Press, Los Alamitos (?), CA
(1991).
I would like to know more about such conference.
Thank You. Marco.
------------------------------
Subject: Stability proofs for recurrent networks
From: steck@spock.wsu.UKans.EDU (jim steck (ME))
Date: Fri, 19 Jul 91 11:51:05 -0500
I am currently looking at the stability issue of syncronous fully
recurrent neural networks.
I am aware of literature (mainly for Hopfield networks) where
the stability issue is adressed using Lyapunov "energy" functions, but
have not seen any publication of other types of approaches. I would
appreciate E-mail regarding articles where this problem is discussed.
Jim Steck
steck@spock.wsu.ukans.edu
------------------------------
End of Neuron Digest [Volume 7 Issue 41]
****************************************
