Copy Link
Add to Bookmark
Report
Neuron Digest Volume 03 Number 09
NEURON Digest Mon Apr 11 09:26:40 CDT 1988 Volume 3 / Issue 9
Today's Topics:
Query on History of NN's
Character font recognition
SUN based NN simulator ???
Carver Mead's book
request for net software
Re: simula
Re: Query on History of NN's
Re simula
Re: display tool for output from neural networks
Stanford Adaptive Networks Colloquium
Seminar Announcement -- Luis ALmeida at GTE
Talk Announcement
Complexity Theory and Hopfield Nets
----------------------------------------------------------------------
Date: 10 Mar 88 07:37:47 GMT
From: Doug Salot <oliveb!felix!dhw68k!doug@ames.arc.nasa.gov>
Subject: Query on History of NN's
I'm curious about what those who are familiar with neural-net literature
consider to be neural-net epochs. What papers are considered seminal?
In a cursory examination of the literature, I'd have to say that the
history goes something like Turing (1936), McCulloch & Pitts (1943),
Hebb (1949), Rosenblatt (1966), Minsky & Papert (1969), and after that
it's not at all clear to me what happens. Grossberg (late '70s)?
Wilshaw? Sutton & Barto? Hopfield?
Would you say Wiener and Cybernetics was a major influence? What about
Leibniz or Shannon?
BTW, has anyone considered using Usenet as a large grained neural network
to which you throw out a question like "what is the meaning of life?" and
watch it converge on a solution?
Thanks in advance for helping me complete this partial match,
- Doug
--
Doug Salot | {trwrb,hplabs}!felix!dhw68k!feedme!doug
CSUF School of Computer Thought | doug@dhw86k.cts.com
"The cobweb behind the Orange Curtain"| If it needs a :-), it isn't funny.
------------------------------
Date: 18 Mar 88 13:34:12 GMT
From: Stan Dembinski <eplrx7!stan@uunet.uu.net>
Subject: Character font recognition
Has anybody been addressing the problem of character font
recognition? I have a vague recollection of a paper having
looked at a limited font set (the Scientific American
magazine font was part of the set). Does anyone know
that ( or other) reference? Neural Net as well as other
more classical approaches are welcome. Thank you in advance,
--
Stan Dembinski | E.I. Dupont Co.
uunet!eplrx7!stan | Engineering Physics Lab
| Wilmington, Delaware 19891
(302) 695-7947 | Mail Stop: E357-311
------------------------------
Date: 23 Feb 88 20:44:00 GMT
From: clyde!burl!codas!novavax!hcx1!garyb@rutgers.edu
Subject: SUN based NN simulator ???
I have seen mention of a SUN based neural-net simulator developed at
Caltech in a recent article from this newsgroup. Could someone
in-the-know elaborate on this. I'm specifically interested in its
functionality, portability to other UNIX platforms, and its
availability status.
I'm also looking for other references to work in area of neural-net
specification languages and mechanisms. I'll summarize and post if
necessary. Thank you for your support.
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
| Gary Barton | ..from the home of the HCX.. |
| Software Development | Harris Computer Systems Division |
+-----------------------------------------+----------------------------------+
| garyb@ssd.harris.com | 2101 W. Cypress Creek Rd. |
| {uunet,cbosgd,mit-eddie}!hcx1!garyb | Ft. Lauderdale. FL 33309 |
| {mtune,gatech}!codas!novavax!hcx1!garyb | (305) 974-1700 |
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
------------------------------
Date: Mon, 21 Mar 88 14:00 EDT
From: DAVIS@blue.sdr.slb.com
Subject: Carver Mead's book
X-Vms-To: MRGATE::M_SDR::IN%"neuron@ti-csl"
Hi. Somewhere, or via someone, I have heard that Carver Mead,
man of optical neural nets, has a book due out called, as I
recall, "Analog VLSI and the brain". Has anyone come across
this, or does anyone know when its due out ? Carver - if you're
out there, please let me know.....
with mnay thanks,
Paul Davis
Schlumberger Cambridge Research,
England.
davis%m_scrvx2@sdr.slb.com
------------------------------
Date: 21 Mar 88 19:05:58 GMT
From: "Wayne D. T. Johnson" <hubcap!ncrcae!ncr-sd!ncrlnk!ncrcce!c10sd3!c10sd1!johnson@gatech.edu>
Subject: request for net software
I have recently read an article on neural nets in the last issue of discovery
(I think). As a software Engineer (AKA Programmer) I would be very
interested if any one out there could direct me to a source of Public
Domain (term used genericly, including such classes as shareware,
freeware, etc.) software for UNIX or an IBM PC/Compatible that could be
useful to a basement experimentor such as I.
I would also like to start a list of basic texts containing information
on nets. Not that some of the information in this group isn't useful
its just that sometimes it goes so far over my head....
If any one would like to contribute any information, please send it to me via
E-mail.
If any one would like a copy of what I receive, send me a self addressed
stamped E-mail envlope and I will try to send it back.
Thanks in advance
Wayne Johnson
------------------------------
Date: 10 Mar 88 13:20:29 GMT
From: Robert Claeson <mcvax!enea!pvab!robert@uunet.uu.net>
Subject: Re: simula
In article <464@sbsvax.UUCP>, ks@sbsvax.UUCP (Kurt Schreiner) writes:
> we are looking for a simula system (compiler or interpreter), preferably
> simula67 which runs under unix (4.3bsd preferred, but others could be hacked)
> or siemens bs2000. a PD version would be most exiting, but hints to lowcost
> custom versions are also welcome.
Sun's Catalyst gives this:
Simula
General Purpose, object-oriented, high-level
programming language.
SIMPROG AB
P.O. Box 26016
S-100 41 Stockholm
Sweden
Tel: +46 8 109912
TLX: 16871
There may be versions for a VAX too.
------------------------------
Date: 11 Mar 88 20:11:36 GMT
From: Olivier Brousse <olivier@boulder.colorado.edu>
Subject: Re: Query on History of NN's
In article <5779@dhw68k.cts.com> doug@dhw68k.cts.com (Doug Salot) writes:
>I'm curious about what those who are familiar with neural-net literature
Roughly speaking, I would say:
Late seventies: Kohonen, on associative memories
" " : Grossberg, on Adaptive Resonance Theory
" " : Barto and Sutton
Early eighties: McClelland, Rumelhart and the PDP group: Back-propagation,
Boltzmann machines, Harmony theory, distributed
representations, cognitive process modeling.
" " : Hopfield, on analogy with physical systems.
Olivier Brousse |
Department of Computer Science | olivier@boulder.colorado.EDU
U. of Colorado, Boulder |
------------------------------
Date: 13 Mar 1988 19:20-EST
From: Tor Sverre Lande <bassen@ifi.uio.no>
Subject: Re simula
If your ever think of using SIMULA-67 on UNIX the only usable
version is the one from Lund, Sweden. We are using it both on
SUN and VAX-en with pretty good performance both on
compilation and execution. That portable stuff (S-PORT) is
totally unusable.
Tor Sverre Lande
Institute of Informatics
University of Oslo
Norway (Simula-land?)
------------------------------
Date: 15 Mar 88 06:36:20 GMT
From: Jeanne Rich <agate!saturn!rich@ucbvax.berkeley.edu>
Subject: Re: display tool for output from neural networks
In article <405@grian.UUCP> liz@grian.UUCP (Liz Allen-Mitchell) writes:
>Does anyone have software that take numbers output from a neural
>network and display them graphically? For example, something that read
>sets of numbers representing the activation levels at a set of nodes
>and displayed them as gray levels in boxes would be great. I am
>working on a sun 3/60 so software that would run on a sun would be best
>though anything would be better than nothing.
>
Yes, The Rochester Simulation Package provides a graphics package with
it, which runs under suntools. You can look at the unit activations,
among other things. The package is fairly flexible, allowing you to
construct any kind of network. I believe their was a recent posting
on who to contact at Rochester. I also believe the cost of the
package is $150.00
Jeanne M. Rich
rich@saturn.ucsc.edu
ucbvax!ucscc!saturn!rich
CIS Board
UCSC
Santa Cruz, CA 95064
(408) 429-4043
------------------------------
Date: Wed, 23 Mar 88 07:50:37 PST
From: Mark Gluck <netlist@psych.stanford.edu>
Subject: Stanford Adaptive Networks Colloquium
Stanford University Interdisciplinary Colloquium Series:
Adaptive Networks and their Applications
March 29th (Tuesday, 3:15pm)
"Casting Neural Networks into Silicon:
there's good news, and there's bad news.
DAN HAMMERSTROM
Computer Science&Engineering
The Oregon Graduate Center
Abstract
--------
Researchers are developing neural-like network models that exhi-
bit a broad range of cognitive behavior. Unfortunately, existing
computer systems are limited in their ability to emulate such
networks efficiently. Consequently, the OGC Cognitive Architec-
ture Project is studying the implementation of massively parallel
architectures for the emulation of a range of very large
connectionist/neural networks. The goal of our project is to
build ultra-large-scale-integrated, silicon-based computing
structures that will be able to emulate connectionist/neural net-
works with thousands of nodes and millions of connections at
rates exceeding that of their biological counterparts. The
manufacturing costs for these systems will be a few thousand dol-
lars. Such ultra-large die size (larger than the traditional 1
square centimeter) is made possible by the inherent fault-
tolerance of the computational model. This talk will explore
some of the problems and potential solutions in developing such
architectures.
. . . .
Format: Tea will be served 15 minutes prior to the talk, outside
the lecture hall. The talks (including discussion) last about
one hour. Following the talk, there will be a reception in
the fourth floor lounge of the Psychology Dept.
Location: Room 380-380W, which can be reached through the lower level
between the Psychology and Mathematical Sciences buildings.
Technical Level: These talks will be technically oriented and are intended
for persons actively working in related areas. They are not intended
for the newcomer seeking general introductory material.
Information: For additional information, contact Mark Gluck
(gluck@psych.stanford.edu) 415-725-2434.
* * *
Co-Sponsored by: Departments of Electrical Engineering (B. Widrow) and
Psychology (D. Rumelhart, M. Gluck), Stanford Univ.
------------------------------
Date: Fri, 25 Mar 88 11:06:53 EST
From: Rich Sutton <rich@gte-labs.csnet>
Subject: Seminar Announcement -- Luis ALmeida at GTE
BACKPROPAGATION IN NONFEEDFORWARD NETWORKS
Luis B. Almeida
INESC, Portugal
10 AM, April 8th, in the GTE Labs Auditorium
The subject of this talk will be the extension of the backpropagation
learning rule to nonfeedforward networks. The classical backpropagation
rule for feedforward networks will first be reviewed, in terms of
operations on networks (linearization and transposition), in order to
better visualize what the solution for nonfeedforward networks might be.
The proof of the validity of this solution will then be given. Next,
stability issues will be addressed. Finally, some experimental results
will be presented.
-----------------------------------------------------------------------
In my opinion, Almeida has done some of the best work on generalizing
the back-propagation learning algorithm to recurrent (cycle containing)
networks (see his paper in the ICNN-87 proceedings). He will be
spending the day at GTE, so there should be ample opportunity for
discussions. Visitors should arrive early and ask for Mary Anne Fox.
GTE Labs is off Rt. 128 at Waltham, west off the Winter St. exit. For
questions contact me (466-4133) or Mary Anne Fox (466-4207).
-Rich Sutton
------------------------------
Date: 11 Mar 88 17:05:35 GMT
From: Kangsuk Lee <siemens!demon!kslee@princeton.edu>
Subject: Talk Announcement
\fBSome Perspectives on Analog Computation\fR
\fIBradley W. Dickinson\fR
Dept. of Electrical Engineering
Princeton University
Princeton, NJ 08544
\fBDate: March 23 (Wed) 10:00 \fR
\fBPlace: Siemens RTL, Princeton NJ \fR
\fBContact: Tom Petsche (609) 734-3392\fR
The term \fIanalog computation\fR evokes numerous associations; the
planimeter, the slide rule, the differential analyzer, and other
physical computing systems are modeled by mathematical equations
with diverse applications. By drawing a careful distinction between
analog and digital computation, it becomes possible to compare the
use of analog computation in the solution of various optimization
problems with the use of digital computation.
Can analog computation be used to solve intractable combinatorial
optimization problems, circumventing the apparent limitations of
digital computation? We will argue that this is unlikely, based on a
simulation paradigm.
Differential equation models associated with $bold { NP }$-complete
problems have been proposed in the literature. These provide an
opportunity to explore and develop interesting complexity issues
related to dynamic systems. A common shortcoming is the failure of the
model to admit a scaling that constrains the solutions to evolve in a
polynomially-bounded hypercube in ``configuration-space'', for a fixed
level of precision of the computation. Analog sorting schemes
provide a simple illustration of this inherent
obstacle to solution of ``numerical'' problems by analog means.
The differential equation models for ``nonnumerical'' problems
can also suffer from this drawback. One topic to be discussed
is the ``neural network'' formulation proposed by
Hopfield and Tank for the solution of the $bold { NP }$-complete
Traveling Salesman Problem with systems described
by differential equations. The scaling problem in this model arises from
the use of highly nonlinear amplifier characteristics.
Time permitting, some possible connections with ergodic theory and
complex dynamics (chaos) will be mentioned. It appears that even
in dynamic systems where scaling problems do not arise, the ability to
do computation may be severely limited.
Kangsuk Lee, Siemens RTL Learning Systems Lab
105 College Road, Princeton, NJ 08540
e-mail: kslee@siemens.com or princeton!siemens!kslee
------------------------------
Date: Sun, 13 Mar 88 12:41:54 EST
From: John Lipscomb <lipscomb@ai.toronto.edu>
Subject: Complexity Theory and Hopfield Nets
Two M.Sc. theses are available by writing or e-mailing me.
They are
The Computational Complexity of
the Stable Configuration Problem
for Connectionist Models
Gail H. Godbeer
September 1987
(latex typesetting)
\begin {abstract}
Connectionist models (CM's) are typically used to perform
constraint-satisfaction searches.
We know that the problem of finding the configuration
that least violates the constraints in a CM is $NP$-hard.
We thus look at the complexity of finding {\em stable} configurations,
or configurations in which all the local constraints are satisfied.
The complexity of finding a stable configuration varies greatly
depending on the type of weights in a CM. We thus classify CM's
according to their weights and examine the parallel and sequential complexity
of this problem for the various classes.
\end {abstract}
AND
On the Computational Complexity of Finding
a Connectionist Model's Stable State Vectors
John Lipscomb
October 1987
\newcommand{\cm}{{\sc cm}}
\newcommand{\ssv}{stable state vector}
\begin{abstract}
Our {\em Connectionist Models\/} (\cm's) are fixed weighted simple graphs
where each node can be in one of two states: {\em on}\/ or {\em off}.
A {\em \ssv\/} for a \cm\ is an assignment of states to the
nodes such that each node is stable.
The stability of a node depends on its state, its weight,
incident edge weights, and neighboring nodes' states.
Our principle results are
\begin{itemize}
\item The decision problem ``Given a directed \cm, does it have a \ssv?''
is NP-complete.
\item The search problem ``Given an undirected \cm, find a \ssv\ for it.''
is P-hard under \NC{1} reductions.
\item The decision problem ``Given an undirected \cm, does it have at
least two different \ssv s?'' is NP-complete.
\item The decision problem ``Given an undirected \cm,
a node $x$, and a state $s$, does the \cm\
have a \ssv\ with node $x$ in state $s$?'' is NP-complete.
\end{itemize}
We consider the complexity of these problems
on more restricted classes of \cm's, and prove that
the problem of finding a \ssv\ can be solved in O($m$) sequential time,
where $m$ is the number of edges, for \cm's with only
positive edges or for bipartite \cm's with only negative edges.
The obvious major question is still open:
``Is there a poly-time algorithm for finding
a \ssv\ for an undirected \cm?''
Also of interest is whether any interesting restricted
class of \cm's has a fast parallel algorithm for finding a \ssv.
\end{abstract}
The theses overlap in parts, but the
different perspective is worthwhile.
John Lipscomb
Dept. of Computer Sc.
University of Toronto
Toronto, Canada
M5S 1A4
(416) 978-4837
home ph: 928-3310
electronic mail:
lipscomb@ai.toronto.edu (CSnet,UUCP,Bitnet)
{uunet,watmath}!ai.toronto.edu!lipscomb
------------------------------
End of NEURON-Digest
********************
