Copy Link
Add to Bookmark
Report
Neuron Digest Volume 04 Number 18
Neuron Digest Friday, 28 Oct 1988 Volume 4 : Issue 18
Today's Topics:
Administrivia^n
Re: Neuron Digest V4 #10
MacBrain
newsletter submission
Rumelhart's generalized cost metric.
Music and PDP (II)
Help with two Companies with N-N products
Re: Help with two Companies with N-N products
theory of itineraries
Wanted: info about GENESIS program
Systolic Architectures for NNs Simulation
Send submissions, questions, address maintenance and requests for old issues to
"neuron-request@hplabs.hp.com" or "{any backbone,uunet}!hplabs!neuron-request"
------------------------------------------------------------
Subject: Administrivia^n
From: "Neuron-Digest Moderator -- Peter Marvit" <neuron@hplms2>
Date: Tue, 01 Oct 88 15:00:08 -0700
[[ It appears, to my embarrassment, that the duplication problem has gotten
worse, not better. I'm almost ready to give up this Digest business... but
not yet. This digest is being sent through yet another route, so once
again, please send me the headers of any duplications.
On another note, this digest begins with a submission which I originally
rejected. After some correspondence with the author, I've decided to include
it and leave you readers to judge the quality of the submission. I've added
some comments, however, since I take editorial privilege.
A note for Keith Loris. I am unable to send any mail directly to you.
Please send headers of this Digest so I can tell the proper route by which
you get this. Otherwise send me your phone number and I'll call you.
Finally, I am interested in hearing from people who are working closer to
the neurophysiological and anatomical side of Artifical Neural nets (i.e,
using neural nets to model "wetware"). Much of the publications that I see
seem to be biased toward the computer science side, though admittedly I
don't comb the biological journals. -PM ]]
------------------------------
Subject: Re: Neuron Digest V4 #10
From: CLAU@A.ISI.EDU
Date: 03 Oct 88 13:39:00 -0400
Subj: Re: Too many neural network meetings
I am of the opinion that there are definitely too too many neural networks
meetings. I have personally heard the same talks (almost identical word by
word) four or five times at different places. With talks given every two
or three weeks, I know there can't be that much research results to report.
There is a practice of reading a few papers, not fully understanding them,
and then slap a few viewgraphs together with some preconceived notions and
present it as a "scientific paper." This kind of practice exploits the
fact that since neural networks research is such a multi-disciplinary
subject that nobody would understand and catch the errors. I believe this
is intellectually dishonest and deplorable. What can be done? I think the
IEEE senior leadership should come out and insist on rigorous review of
submitted papers before presentation or publication. Unfortunately some of
the senior people are guilty of this practice themselves. I just hope that
in a democracy the people will eventually recognize this for what it is,
and get down to the serious business of doing good neural networks research
which promises so much potential both in terms of understanding brain
function and building high performance processors.
[[ Connectionism, aka Neural Nets, while a sub-discipline of itself, can
better be described as a new modeling paradigm and a new set of vocabulary
to apply to problems. The quality of publications has certainly been
uneven over the past few years; while the situation is improving, I see
this as a sign of an immature field and not as wanton opportunism or even
as "intellectually dishonest." The variance in quality can be shown in
nearly every field, I contend.
Further, Connectionism has the great virtue, which may underly some of the
early excesses, of offering immediate results in many areas to people who
may not otherwise have the background for other types of computer modeling.
It allows old problems to be descibed in a very different light.
I also see that many people are working on similar problems but do not know
of each others work. The routes of communication are still only beginning
to be established, and the field is still sorting itself into
subspecialties. I am not ready to condemn the participants wholesale in
the natural progression of scientific knowledge.
I would counter that there are many papers, partially because there are so
many people working in the new field and so much is being discovered.
There are a few "stars" whose names recur, and perhaps this is what the
writer rails against; surely one person can produce only so much, even if
he or she is invited to speak only weeks apart. Yet, papers on the same
topic by the same person shows the incremental increase in knowledge; not
all papers can be revolutionary "blockbusters." But also of the hundreds
or thousands of working and publishing?
Finally, Connectionism, like all branches of science, is composed of
*people* working on problems -- people with their own weaknesses, foibles,
egos, and visions. The field advances no faster than the people in it. I'm
confidant that most conferences and journals feel their review process is
rigorous, but it is still subject to the personal characteristics, politics
and biases of the reviewers. Science is neither a democracy nor
objectively black and white; for years, the idea of plate tectonics was
dismissed by the majority.
I am all for maintaining the highest scientific standards as possible. I
do not expect human perfection. Errors will happen. You, as astute
observers, will ultimately catch them. If a person willfully presents
erroneous information, he or she loses credibility -- the only thing a
scientist has. It is therefore *your* job, as scientists and engineers, to
cast a skeptical eye on every paper, follow those which seem most
promising, and refute those which are astray. Use caution, however, when
judging the people behind them. -PM
------------------------------
Subject: MacBrain
From: joe@epistemi.ed.ac.uk (Joe Levy)
Organization: Centre for Cognitive Science, Edinburgh, UK
Date: 17 Oct 88 14:20:48 +0000
Has anyone out there used MacBrain 1.2, the neural net simulator
for the Macintosh? I'm especially interested in comments on its
suitability for
(a) Teaching and demonstration purposes
(b) Research
Have there been any reviews of this program?
Has anyone in the UK or Europe managed to get a copy?
Joe Levy Phone: +44 31 667 1011 x6232 | University of Edinburgh
UUCP: ...!uunet!mcvax!ukc!its63b!epistemi!joe | Centre for Cognitive Science
ARPA: joe%epistemi.ed.ac.uk@nss.cs.ucl.ac.uk | 2 Buccleuch Place
JANET: joe@uk.ac.ed.epistemi | Edinburgh EH8 9LW Scotland
------------------------------
Subject: newsletter submission
From: Derek Smith <DSMITH@vdle20.csc.ti.com>
Date: Mon, 17 Oct 88 15:15:58 -0500
IEEE Conference on "NEURAL INFORMATION PROCESSING SYSTEMS -
NATURAL AND SYNTHETIC"
I am interested in sharing insights, results, and ideas on Pentti Kanerva's
SPARSE DISTRIBUTED MEMORY and David Marr's THEORY OF THE CEREBELLAR
CORTEX.
If you are too, then maybe we can meet during the NIPS conferenece or
at the post-meeting workshop. Send email.
Derek Smith.
[[ Kanerva has also written the book "Sparse Distributed Memory" to be
published in December by MIT Press. Check to see if there might be an
advance copy at NIPS. -PM]]
------------------------------
Subject: Rumelhart's generalized cost metric.
From: Luis Borges de Almeida <mcvax!inesc!lba@uunet.uu.net>
Date: Tue, 18 Oct 88 10:15:02 +0000
I did not attend Rumelhart's talk, but I would like to suggest the
following modifications of the formulas posted by Matt Heffron (at
least they make sense to me):
cost = L*(<the-regular-sum-squared-error>) +
-----
\
(1 - L) * ((K1 * > <cost-per-weight>) +
/
-----
i,j
-----
\
(K2 * > <cost-per-unit>))
/
-----
i
( ) 2
( w )
( ij )
<cost-per-weight> = -----------------
( ) 2
1 + ( w )
( ij )
and
----
\ ( ) 2
> ( w )
/ ( ik )
----
k
<cost-per-unit> = -------------------
----
\ ( ) 2
1 + > ( w )
/ ( ik )
----
k
Luis B. Almeida
uucp: ...mcvax!inesc!lba
------------------------------
Subject: Music and PDP (II)
From: MUSICO%BGERUG51.BITNET@CUNYVM.CUNY.EDU
Date: Wed, 19 Oct 88 15:23:00 +0100
There are other papers on Music and PDP. See :
J. Bharucha : "Neural Net Modeling of Music",
M. Leman : "Sequential (Musical) Information Processing
with PDP-Networks",
B. Vercoe : "Hearing Polyphonic Music with the Connection Machine"
in : Proceedings of the first workshop on Music and AI.
Work is also done by C. Lischka (See the Proceedings of the ICMC 87 and the
Arbeidspapiere der GMD). I have two other related papers ("Neural Net-
works in Music Research" and "Massive Parallel Computer Methods in Music
Research"). I currently revise these papers and they will very soon be
available.
I guess that many other people are working in the same direction.
Marc Leman
University of Ghent
Institute for Psychoacoustics and Electronic Music
Blandijnberg 2
B-9000 GHENT
Belgium
------------------------------
Subject: Help with two Companies with N-N products
From: darylm@illian.UUCP (Daryl V. McDaniel)
Organization: Micronetics, Aloha, Oregon
Date: 19 Oct 88 15:46:00 +0000
I am trying to get some information on two companies that are supposed to
be producing products based upon neural networks.
The first company is "Thunderstone/Expansion Programs International, Inc.
(EPI). They have a text retrieval system which they call "Metamorph" that
can "find and match concepts in unindexed text" by using "morphemic"
search. "In an IBM/AT microcomputer, Metamorph can search unindexed ASCII
text at the rate of 500,000 characters per second.
The only Information I have is:
Metamorph is based upon neural-network technology
CEO is Michael Pincus
The company has seven employees
They have had products on the market since 1985 (or before?)
EPI used Government IR&D funds to develop Metamorph
EPI is based in Ohio
The National Security Agency now buys all its ceramic packages
from Kyocera, a Japanese company.
Is this company real? Do their products work? DARPA supposedly thinks
that they are years ahead of the general state of the technology. Are
they? Has anyone evaluated Metamorph?
The other company is "Extrema Systems, International", (XSI), of Reston,
Virginia. (I called information, they've never heard of them)
To paraphrase XSI's co-founder and VP of R&D, Dr. Arle Visser: Extremas are
the extreme points of waveforms. By measuring these, rather than the
frequency or amplitude fo signals received, a final "clean" signal can be
produced more efficiently. The signal remains "in context" with its
background at all times. All digital methods of signal processing are
inappropriate to the real world. The XSI system closely fits the way we
think the brain carries out processing of perceptual stimuli. First, the
brain processes phase components of waveforms. Second, input stimuli
consists of variations of amplitude levels over a wide dynamic range.
Amplitude detection, therefore, is bit-intensive. The Extrema device's
architecture contains 3 components: a differentiator, a broadband infinite
clipper, and an integrator. The 3 together process a signal by measuring
its "extremas," identifying the number of times these pass above or below a
zero point (or center frequency) on the waveband and then integrating these
into a final output.
As part of the development work, an Extrema Coding Signal Processor (XCSP,
the proprietary "generic" product of the company) has been connected to the
inner ear of a deaf subject, feeding signals directly to the auditory
nerve. In one case, the subject, in experiments carried out at Stanford
University, CA, reported it was possible to "hear" the signal being fed in
- -- discriminate actual words. "Regular" electrical impulses fed to the
nerve left no impression, other than that "something was happening, to the
subject.
XSI considers its real market in supplying its equipment to larger
corporations. "At this stage we're component manufacturers, so we would
supply to systems houses," says XSI president J. Steven Lovink. Because
these patents are applied for on a world wide basis, he sees no problem
with technology transfer abroad, and indicates that a NATO army is already
a customer for XCSPs.
Again, is this company real? Has anyone seen their work or papers by them?
I would appreciate hearing any comments about either company, their
products, or the technologies they represent. These are two of only three
companies that I have heard of that are actually marketing products based
upon neural-network technology. I would also like to hear of others.
Thank you very much, (My appologies for typo's)
_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-_-
Daryl V. McDaniel .........sun!nosun!illian!darylm
Micronetics USENET: ...tektronix!nosun!illian!darylm
4730 S.W. 182nd Ave. TELEX: WUI 6972206
Aloha, OR 97007 PHONE: (503) 224-7056
------------------------------
Subject: Re: Help with two Companies with N-N products
From: newmen@nprdc.arpa (Raye Newmen)
Organization: Navy Personnel R&D Center, San Diego
Date: 21 Oct 88 16:58:56 +0000
>The first company is "Thunderstone/Expansion Programs International, Inc.
>(EPI). They have a text retrieval system which they call "Metamorph" that
>[...] Has anyone evaluated Metamorph?
Metamorph, a natural language free-text analysis and retrieval system, is
a product of THUNDERSTONE, Expansion Programs International, Inc.
Post Office Box 839
Chesterland, Ohio 44026
(216) 449-6104
We are currently evaluating the MSDOS version. It seems quite powerful if
you extened the lexecon for the specialized vocabulary of your application
domain. I understand that the UNIX version (MM 3.0) is greatly enhansed,
faster, and modular so the "Metamorph engine" can be integrated with your
own application interface. We have not had an opportunity to test this
product.
>The other company is "Extrema Systems, International", (XSI), of Reston,
>Virginia. (I called information, they've never heard of them)
>[....]
>Again, is this company real? Has anyone seen their work or papers by them?
I did visit XSI, saw a demo and got some handouts. I have not been able
to evaluate their approach or products. What I did see of their
technology was very interesting. They are for real. You can contact them
at:
Extrema Systems International Corporation
10805 Park Ridge Blvd., Suite 130
Reston, VA 22091
(703) 648-3181
I hope this helps,
Raye Newmen newmen@nprdc.arpa
------------------------------
Subject: theory of itineraries
From: <IAOM100%INDYVAX.BITNET@CORNELLC.CCS.CORNELL.EDU>
Date: Wed, 19 Oct 88 11:39:00 -0500
Is there anyone out there who knows about the theory of
itineraries as applied to cellular automata. Also, is there anyone doing
convergence proofs on ANNs, and what mathematical tools are they using. I
have been using Lyapunov functions as detailed by E. Goles and G. Vichniac.
I would greatly appreciate any help.
Masud Cader
(IAOM100@INDYVAX) ---BITNET.
------------------------------
Subject: Wanted: info about GENESIS program
From: muffy@violet.berkeley.edu
Organization: University of California, Berkeley
Date: 21 Oct 88 21:52:06 +0000
I would like some information about the GENESIS neural simulation program
being developed at Cal Tech. A friend used it at the Woods Hole course on
Neural Computation but I have not been able to get a hold of him to find
out the details of what the program can do. Does anyone know about this
program?
Kathy Murphy
muffy@violet.berkeley.edu
------------------------------
Subject: Systolic Architectures for NNs Simulation
From: Dario Ringach <dario%TECHUNIX.BITNET@CUNYVM.CUNY.EDU>
Date: Tue, 25 Oct 88 10:15:05 +0200
I'm interested in reconfigurable/programmable systolic architectures for
NNs' simulation. I heard a rumor that there was a talk on this subject at
the second INNS meeting in Boston last September... Does anyone know who
actually gave it?
I'll also appreciate any other references on this topic.
Thanks in advance.
Dario.
BITNET: dario@techunix
------------------------------
End of Neurons Digest
*********************
