Copy Link
Add to Bookmark
Report
Neuron Digest Volume 10 Number 15
Neuron Digest Friday, 6 Nov 1992 Volume 10 : Issue 15
Today's Topics:
New version of Learning Vector Quantization PD program package
Info on intelligent agents?
Economics and Neural Nets bibliography, addendum
Effectiveness of the latest ANNSs
Production scheduling systems?
non-linear dynamical modelling?
Modeling question
Job at Booz, Allen & Hamilton
Request for advice - sound localization
Algorithms for masssivley parallel machines?
Postdocs at Rockefellar
Send submissions, questions, address maintenance, and requests for old
issues to "neuron-request@cattell.psych.upenn.edu". The ftp archives are
available from cattell.psych.upenn.edu (130.91.68.31). Back issues
requested by mail will eventually be sent, but may take a while.
----------------------------------------------------------------------
Subject: New version of Learning Vector Quantization PD program package
From: lvq@cochlea.hut.fi (LVQ_PAK)
Date: Sun, 11 Oct 92 10:57:21 +0700
************************************************************************
* *
* LVQ_PAK *
* *
* The *
* *
* Learning Vector Quantization *
* *
* Program Package *
* *
* Version 2.1 (October 9, 1992) *
* *
* Prepared by the *
* LVQ Programming Team of the *
* Helsinki University of Technology *
* Laboratory of Computer and Information Science *
* Rakentajanaukio 2 C, SF-02150 Espoo *
* FINLAND *
* *
* Copyright (c) 1991,1992 *
* *
************************************************************************
Public-domain programs for Learning Vector Quantization (LVQ)
algorithms are available via anonymous FTP on the Internet.
"What is LVQ?", you may ask --- See the following reference, then:
Teuvo Kohonen. The self-organizing map. Proceedings of the IEEE,
78(9):1464-1480, 1990.
In short, LVQ is a group of methods applicable to statistical
pattern recognition, in which the classes are described by a
relatively small number of codebook vectors, properly placed
within each class zone such that the decision borders are
approximated by the nearest-neighbor rule. Unlike in normal
k-nearest-neighbor (k-nn) classification, the original samples
are not used as codebook vectors, but they tune the latter.
LVQ is concerned with the optimal placement of these codebook
vectors into class zones.
This package contains all the programs necessary for the correct
application of certain LVQ algorithms in an arbitrary statistical
classification or pattern recognition task. To this package three
options for the algorithms, the LVQ1, the LVQ2.1 and the LVQ3,
have been selected.
This code is distributed without charge on an "as is" basis.
There is no warranty of any kind by the authors or by Helsinki
University of Technology.
In the implementation of the LVQ programs we have tried to use as
simple code as possible. Therefore the programs are supposed to
compile in various machines without any specific modifications made on
the code. All programs have been written in ANSI C. The programs are
available in two archive formats, one for the UNIX-environment, the
other for MS-DOS. Both archives contain exactly the same files.
These files can be accessed via FTP as follows:
1. Create an FTP connection from wherever you are to machine
"cochlea.hut.fi". The internet address of this machine is
130.233.168.48, for those who need it.
2. Log in as user "anonymous" with your own e-mail address as password.
3. Change remote directory to "/pub/lvq_pak".
4. At this point FTP should be able to get a listing of files in this
directory with DIR and fetch the ones you want with GET. (The exact
FTP commands you use depend on your local FTP program.) Remember
to use the binary transfer mode for compressed files.
The lvq_pak program package includes the following files:
- Documentation:
README short description of the package
and installation instructions
lvq_doc.ps documentation in (c) PostScript format
lvq_doc.ps.Z same as above but compressed
lvq_doc.txt documentation in ASCII format
- Source file archives (which contain the documentation, too):
lvq_p2r1.exe Self-extracting MS-DOS archive file
lvq_pak-2.1.tar UNIX tape archive file
lvq_pak-2.1.tar.Z same as above but compressed
An example of FTP access is given below
unix> ftp cochlea.hut.fi (or 130.233.168.48)
Name: anonymous
Password: <your email address>
ftp> cd /pub/lvq_pak
ftp> binary
ftp> get lvq_pak-2.1.tar.Z
ftp> quit
unix> uncompress lvq_pak-2.1.tar.Z
unix> tar xvfo lvq_pak-2.1.tar
See file README for further installation instructions.
All comments concerning this package should be
addressed to lvq@cochlea.hut.fi.
************************************************************************
------------------------------
Subject: Info on intelligent agents?
From: Nick Vriend<VRIEND@IFIIUE.FI.CNR.IT>
Date: Mon, 12 Oct 92 16:16:48
Nick Vriend
European University Institute
C.P. 2330
50100 Firenze Ferrovia
Italy
EARN/Bitnet: <VRIEND@IFIIUE.FI.CNR.IT>
As a PhD student of economics at the European University Institute in
Florence (Italy), finishing a thesis on 'Decentralized Trade', I am
interested in getting contact with people who are working on the
following topic: DECENTRALIZED TRADE WITH ARTIFICIALLY INTELLIGENT
AGENTS. Basic characteristic of decentralized economies is that each
individual agent has a very limited knowledge of his relevant
environment. Each agent acts and observes his outcomes in the market
(which depend on the actions of the other participants). Thus, each
individual agents learns independently, using only a success measure of
his own actual performance (e.g. profits, utility).
At the moment I am applying Classifier Systems and Genetic Algorithms to
model the learning process of each individual agent, but (given the
mentioned inherent problem of misspecification in decentralized
economies) Neural Networks seem very promising. However, application of
Neural Networks appears more complex, as in a decentralized economy
nobody would be able to tell each agent what his "target" or "correct"
decision would have been. Therefore, the machines have to learn
unsupervised (as in e.g. Barto, Sutton & Anderson (1983): Neuronlike
Adaptive Elements That Can Solve Difficult Learning Control Problems.
IEEE Transactions on Systems, Man, and Cybernetics, 13). Hence, the
topic I am interested in might be restated as: REINFORCEMENT LEARNING BY
INTERACTING MACHINES.
------------------------------
Subject: Economics and Neural Nets bibliography, addendum
From: Duarte Trigueiros <dmt@sara.inesc.pt>
Date: Wed, 14 Oct 92 11:27:46 -0200
In addition to Paul Refenes list, I would like to mention mine and Bob's
paper on the automatic forming of ratios as internal representations of
the MLP. This paper shows that the problem of discovering the appropriate
ratios for performing a given task in financial statement analysis can be
be simplified by using some specific training schemes in an MLP.
@inproceedings( xxx ,
author = "Trigueiros, D. and Berry, R.",
title = "The Application of Neural Network Based Methods to the
Extraction of knowledge From Accounting Reports",
Booktitle = "Organisational Systems and Technology: Proceedings of the
$24^{th}$ Hawaii International Conference on System
Sciences",
Year = 1991,
Pages = "136-146",
Publisher = "IEEE Computer Society Press, Los Alamitos, (CA) US.",
Editor = "Nunamaker, E. and Sprague, R.")
I also noticed that Paul didn't mention Utans and Moody's "Senlecting
Neural Network Architectures via the Prediction Risk: An Application to
Corporate Bond Rating Prediction" (1991), which has been published
somewhere and has, or had, a version in the neuroprose archive as
utans.bondrating.ps.Z .This paper is especially recommended, as the early
literature on financial applications of NNs didn't care too much with
things like cross-validation. The achievements, of course, were
appallingly brilliant.
Finally, I gathered from Paul's list of articles, that there is a book of
readings entitled "Neural Network Applications in Investment and
Finance". Paul is the author of an article in chapter 27. The remaining
twenty six or so chapters can eventually contain interesting stuff for
completing this search.
When the original request for references appeared in another list I
answered to it. So, I must apologise for mentioning our reference again
here. I did it, as Paul list of references could give the impression,
despite him, of being an attempt to be extensive.
---------------------------------------------------
Duarte Trigueiros,
INESC, R. Alves Redol 9, 2. 1000 Lisbon, Portugal
e-mail: dmt@sara.inesc.pt FAX +351 (1) 7964710
---------------------------------------------------
------------------------------
Subject: Effectiveness of the latest ANNSs
From: Fernando Passold <EEL3FPS%BRUFSC.bitnet@UICVM.UIC.EDU>
Organization: Universidade Federal de Santa Catarina/BRASIL
Date: Thu, 15 Oct 92 14:39:35 -0300
I would like to begin a little discussion questioning the effectiveness
of the latest Artificial Neural Networks Simulators (ANNSs).
A majority of the ANNSs do a serial computation trying to emulate the
process of the Natural (or biological) Neural Networks (NNNs).
I would like to drive the attentions about the fact that the computation
is doing serialy, or better, even in more improved ANNSs that profit parallel
and/or concurrent processing, what is computing it's one synapse at time or
in blocks (packets), different that occurs with NNNs. In a NNN, various
potentials of activations of neurons (membrane voltage changes) are evoluate
and processing at the same time (involving different latencies), and not the
outcome of each synapse from time to time (like in conventionals ANNSs).
The question that I would arise follows: Imagine that one of the neurons
of an especific NNN suddenly presents a bigger latency in its response,
compared with its neighbourhoods. Will it be that this 'failed' neuron, do
not carry this network to a completely different result than it would be
expected ? Will it be that the activation's timing (spike timing)
between neurons of an especific network do not deserve too more attention
than the mere emulating of the majority of ours latest ANNSs, even with
parallel and/or concurrent processing ?
Would not this synchronism (above mentioned) be responsible for our
primitive intuitive notion of velocity and time (epistemological talking) ?
Maybe this discussion would be of greater interest for the people of
Construtivist IA (or Construtivist Connectionist IA).
I would be glad in order to receive opinious and/or even outcomings
form researchs (preferential via neuron-digest list) with this in mind
(including 'neural-boards' using Analog implementations, DSPs inmplementations,
neuron-chips or transputers outcomings).
Fernando Passold
(Master degree student)
Biomedical Engineering Group
Santa Catarina Federal University
BRAZIL
E-mail: EEL3FPS@BRUFSC.BITNET
------------------------------
Subject: Production scheduling systems?
From: shim@educ.emse.fr
Date: 15 Oct 92 20:03:48 +0000
Can anyone suggest some references or who have worked to apply on the
production scheduling systems with neural networks. And, could someome
forward me Mr. Yoon-Pin Simon Foo's email address(I think he is(or was)
in Univ. South Carolina.).
Thank in advance.
------------------------------
Subject: non-linear dynamical modelling?
From: ABDEMOHA@th.isu.edu
Organization: Idaho State University
Date: 15 Oct 92 20:42:39 -0700
Dear Sir
I would like to inquire about the use of neural networks in modelling
non-linear dynamical systems. This may also include the ability of
the neural networks to approximate systems governed by a pre-known
partial differential equations?.
Mohamed A. Abdel-Rahman
------------------------------
Subject: Modeling question
From: ABDEMOHA@th.isu.edu
Organization: Idaho State University
Date: 15 Oct 92 20:46:33 -0700
Dear Sir:
I was trying to use the backpropagation neural network to approximate
a certain function. I found out that the resulting network could
approxiamte the high value output points more than the low value
points. I think thos maybe dur to that the error surface is an
absolute one. Have there been any trials to construct a relative
error surface (i.e. |(Fa(w) - Fd(w))/Fa(w)|).
Mohamed A. AbdelRahman
------------------------------
Subject: Job at Booz, Allen & Hamilton
From: brettle@picard.ads.com (Dean Brettle)
Date: Mon, 19 Oct 92 18:15:51 -0500
NEURAL NETWORK DEVELOPERS
Booz, Allen & Hamilton, a world leader in using technology to solve
problems for government and industry, has immediate openings for
experienced neural network developers in our Advanced Computational
Technologies group.
If chosen, you will help develop, implement and evaluate neural
network architectures for image and signal processing, automatic
target recognition, parallel processing, speech processing, and
communications. This will involve client-funded work as well as
internal research and development. To qualify, you must have
experience in neural network theory, implementation, and testing;
C/UNIX/X11; and parallel processing experience is a plus.
Equal Opportunity Employer. U.S. citizenship may be required.
Candidates selected will be subject to a background investigation and
must meet eligibility requirements for access to classified
information.
ENTRY-LEVEL CANDIDATES should have a BS or MS degree in either
computer science, applied mathematics, or some closely related
discipline and experience implementing neural network paradigms.
MID-LEVEL CANDIDATES should have a BS or MS degree in either computer
science, applied mathematics, or some closely related discipline with
>3 years experience. Candidates must possess a working knowledge of
popular neural network models and experience implementing several
neural network paradigms.
SENIOR-LEVEL CANDIDATES should have an MS or Ph.D. degree in either
computer science, mathematics, computational neuroscience, electrical
engineering or some closely related discipline. Published work and/or
presentations in the neural network field are highly desirable.
Experience applying neural network technology to real-world problems
and in developing neural network programs (including marketing,
proposal writing, and technical & contractual management) is required.
Booz, Allen offers a competitive salary, excellent benefits package,
challenging work environment and ample opportunities to advance your
career. Please send a resume to Dean Brettle either by email to
brettle@picard.ads.com, fax to (703)902-3663, or surface mail to Booz,
Allen & Hamilton Inc., 8283 Greensboro Drive, Room 594, McLean, VA
22102.
------------------------------
Subject: Request for advice - sound localization
From: net@sdl.psych.wright.edu (Mr. Happy Net)
Date: Tue, 20 Oct 92 02:17:24 -0500
Dear Sir,
At Wright State University, we are working on developing an
artificial neural net model of human sound localization. One of our
objectives has been to show that ANN's adhere to the Duplex Theory of
Localization in that they make use of high frequency intensity cues over
low frequency intensity cues, and low frequency temporal cues over high
frequency temporal cues. We have chosen to use the backpropagation
algorithm distributed in the NeuralShell package available from Ohio
State University (anonymous ftp quanta.eng.ohio-state.edu).
One of our approachs has been to train ANN's with low, mid, or high band
filtered signals. A problem with this is that in humans, our
"net" learns to deal with broad band signals by selecting which portions
of the signal to base judegments on. On the other hand, if we train an
ANN with broad band signals, we would like to uncover which portions of
the input spectrum are most heavily affecting the ANN's decisions. This
is difficult to do because we cannot merely zero out portions of the
input spectrum and test the ANN's performance, as such provides false
cues indicating the signal is comming from either directly in front of
or behind the head. I would greatly appreciate any suggestions on how
to analyze the "weighting" the net gives to different portions of its input.
Jim Janko
net@sdl.psych.wright.edu
------------------------------
Subject: Algorithms for masssivley parallel machines?
From: "Rogene Eichler" <eichler@pi18.arc.umn.edu>
Date: Wed, 21 Oct 92 12:13:49 -0600
I am looking for references describing optimization algorithms for
backprop type networks on either the CM-200 or CM-5. i.e. What algorithms
best exploit the massive parallelism?
Thanks!
- Rogene
eichler@ahpcrc.umn.edu
------------------------------
Subject: Postdocs at Rockefellar
From: Zhaoping Li <zl%venezia.ROCKEFELLER.EDU@ROCKVAX.ROCKEFELLER.EDU>
Date: Thu, 22 Oct 92 11:53:40 -0500
ROCKEFELLER UNIVERSITY
anticipates the opening of one or two positions in Computational
Neuroscience Laboratory. The positions are at the postdoctoral level,
and are for one year, renewable to two, starting in September 1993.
The focus of the research in the lab is on understanding the
computational principles of the nervous system, especially
the sensory pathways. It involves analytical and computational approaches
with strong emphasis on connections with real neurobiology. Members
of the lab include J. Atick, Z. Li, K. Obermayer, N. Redlich, and
P. Penev. The lab also maintains strong interactions with other labs at
Rockefeller University, including the Gilbert, Wiesel, and the biophysics
labs.
Interested candidates should submit a C.V. and arrange to have three
letters of recommendation sent to
Prof. Joseph J. Atick
Head, computational neuroscience lab
The Rockefeller University
1230 York Avenue
New York, NY 10021 USA
The Rockefeller University is an affirmative action/equal opportunity
employer, and welcomes applications from women and minority candidates.
------------------------------
End of Neuron Digest [Volume 10 Issue 15]
*****************************************
