neuron-request@HPLMS2.HPL.HP.COM ("Neuron-Digest Moderator Peter Marvit") (07/14/90)
Neuron Digest Friday, 13 Jul 1990 Volume 6 : Issue 43 Today's Topics: Re: Neuron Digest V6 #42 Re: Neuron Digest V6 #42 neural inhibition Re: A questin about neural inhibition Kohonen's network Protein analysis using ANN References Answer to K. Morse how our genes give us brains Anybody in Poland working in NN Cohen, Dunbar, & McClelland already published? job opening Evolving Networks - New TR Tech Reports Available Report on Non Linear Optimization Call for Participation in Connectionist Natural Language Processing Apologies on Call for Paticipation Technology Transfer Mailing List 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: Re: Neuron Digest V6 #42 From: "Meyer E. Nigri" <M.Nigri@Cs.Ucl.AC.UK> Date: Tue, 10 Jul 90 10:04:50 +0100 David McKee writes: <Digressing a moment, I have been thinking about creating a standard for a <software simulation/hardware description language for neural nets (I work <with simulation languages and HDL's , specifically VHDL). I would like to inform you that such an idea has already been done in the PYGMALION project. Meyer. +--------------------------+-----------------------------------------------+ |Meyer Elias Nigri | JANET:mnigri@uk.ac.ucl.cs | |Dept. of Computer Science | BITNET:mnigri%uk.ac.ucl.cs@UKACRL | |University College London |Internet:mnigri%cs.ucl.ac.uk@nsfnet-relay.ac.uk| |Gower Street | ARPANet:mnigri@cs.ucl.ac.uk | |London WC1E 6BT | UUCP:...!mcvax!ukc!ucl-cs!mnigri | +--------------------------+-------------------------+---------------------+ |Tel: +44 (071)-387-7050 | Fax: +44 (071)-387-1397 | Telex: 28722 | | ext. 3701 | | | +--------------------------+-------------------------+---------------------+ [[ Editor's Note: Perhaps someone would care to explain what the PYGMALION project is or was? -PM ]] ------------------------------ Subject: Re: Neuron Digest V6 #42 From: J. P. Letellier <jp@radar.nrl.navy.mil> Date: Tue, 10 Jul 90 17:50:02 -0400 [[ In previous Digest, "DAVE MCKEE" <mckee@tisss.radc.af.mil> writes: ]] > I would like to pose a question to the list about inhibiting neurons. Dave, Why don't you try your simulations in VHDL? (-: It is optimized for interconnects and duplication of identical parts. That makes it only necessary to embed whatever non-linear function you intend to demonstrate inside the cell. You could actually build a library of various cells, and compare (or even mix) their actions on different functions and problems. Since VHDL is event driven, it should be optimum for this task!! Sounds like fun. If you really don't want to work with VHDL, then I would suggest doing your simulation language in either Ada or C++. In either case, that would allow you to focus better on the objects and their actions. C may be powerful, but you easily get lost in the code and miss the intuition you are trying to establish. jp ------------------------------ Subject: neural inhibition From: tony@helmholtz.sdsc.edu (Tony Bell) Date: Tue, 10 Jul 90 23:50:06 +0100 Dave McKee's question about inhibition that can veto firing is an interesting one. What he is talking about is a non-linear form of inhibition called 'shunting' or 'silent' inhibition. This has been postulated to be a component in direction selectivity in the retinal ganglion cell by Christof Koch [1] though Rodney Douglas et al cast doubt on this being important in visual cortex [2]. Anatomically, though, it is interesting that supposedly inhibitory cortical synapses tend to dominate on the cell body, thick dendrites and the start of the axon, where they can have maximal effect on vetoing or at least modulating the results of excitatory input. I agree with Dave that this is the kind of thing that should be in the connectionist repetoire, especially if we want to take clues from the brain. You can look up my attempt in [3] if you like. It includes a learning algorithm able to deal with non- linearities at synapse or dendritic branchpoint - called Artificial Dendritic Learning. Unfortunately it's static. If anyone is interested, 2 good books approach these biological issues from an introductory (but very detailed) viewpoint [4], and from a computational [5] viewpoint. Tony Bell Refs. [1] Koch et al, Retinal Ganglion Cells: a functional interpretation of dendritic morphology. Phil. Trans. R. Soc. London [Biol] 298: 227-264. [2] Douglas et al, Nature 332:642-644 (1988) [3] Bell T, Higher-order learning in 'Artificial Dendritic Trees', in Touretzky (ed), Advances in Neural Information Processing Systems 2 (1990) [4] The Synaptic Organisation of the Brain, 3rd edition, ed. Gordon Shepherd, 1990. Now in paperback (~$30), Oxford Univ. Press [5] Methods in Neuronal Modeling, eds, Koch C & Segev I, MIT press, 1989 ------------------------------ Subject: Re: A questin about neural inhibition From: Jonathan Delatizky <delatizk@BBN.COM> Date: Wed, 11 Jul 90 10:21:38 -0400 David McKee asks some interesting questions about inhibitory mechanisms in the dendritic fields of real neurons. I'm no longer involved in work on these topics, so there may be newer work about which I'm unaware. For the same reason, I won't be able to give precise citations. Wilfred Rall developed an enormous body of material on models of cable conduction in dendritic fields. Both excitatory and inhibitory inputs were supported in his methodology. The models are detailed, so that the computational complexity of solutions for realistic spatial configurations takes a lot of CPU power. In addition to papers in the biophysical and neurobiological literature, he published at least one monograph on the subject. The idea that some forms of inhibibitory inputs may act to short-circuit a region of the dendritic tree is also well established, though I can give you no names or citations to back this up. Finally, Jerry Lettvin's group at MIT investigated the influence of the configuration of axon terminal branches and their prior history of activation on failure of conduction at axonal terminal branch points. Not quite the same mechanism, and a very different location, but also relevant to the general nature of the query. There is no question, of course, that the "neurons" included in almost all so called neural network models ignore the complex computations that are certainly performed by subthreshold interactions in dendritic fields. This does not invalidate the results of such work, rather suggesting that the field has made some poor terminological choices. ------------------------------ Subject: Kohonen's network From: JJ Merelo <jmerelo@ugr.es> Date: 06 Jul 90 12:49:00 +0200 I am working on Kohonen's network. I was using previously a 15 input, 8x8 output layer network, with parameters stated in Kohonen's 1984 paper ( also in Aleksander book ). Y switched to a 16 input, 9x9 layer output, with the same parameters, and everythi n g got screwed up. Should I use the same parameters? How should them be changed? What do they depend on? Please, somebody answer, or i'll cast my Sun ws thru the window JJ ------------------------------ From: BRUNAK@nbivax.nbi.dk Date: Sun, 08 Jul 90 17:02:00 +0200 Subject: Protein analysis using ANN Title: ``Analysis of the Secondary Structure of the Human Immunodefieciency Virus (HIV) Proteins p17, gp120, and gp41 by Computer Modeling Based on Neural Network Methods'' H. Andreassen, H. Bohr, J. Bohr, S. Brunak, T. Bugge, R.M.J. Cotterill, C. Jacobsen, P. Kusk, B. Lautrup, S.B. Petersen, T. Saermark, and K. Ulrich, in Journal of Acquired Immune Deficiency Syndromes (AIDS), vol. 3, 615-622, 1990. ------------------------------ Subject: references From: demelerb@bionette.CGRB.ORST.EDU (Borries Demeler - Biochem) Date: Tue, 10 Jul 90 12:54:52 -0700 Hi, I would like to find out if neural networks have been, in any way, used for sequence analysis of nucleic acids, i.e., RNA and DNA. If you have some information or references, I would greatly appreciate if you could share them with me by e-mailing to: demelerb@bionette.cgrb.orst.edu (Borries Demeler, Dept. of Biochemistry and Biophysics, Oregon State University, Corvallis, Or. 97331-6503) Thank you very much, -Borries- ------------------------------ Subject: Answer to K. Morse From: JJ Merelo <jmerelo@ugr.es> Date: 11 Jul 90 13:19:00 +0200 As a test-answer to your last question, I think ( and it comes from "The magic loom", by R. Jastrow ) that the brain structure comes from the evolutionary past of man. The brain is composed of several layers, the inner corresponding to the further past ( as reptiles ) and the outer to the present ( as men ). Thus, we would have a reptilian brain layer ( that would roughly correspond to the cerebellum ), a mammalian brain layer ( to some part of the brain, I don't quite remember ) and a outer "human" layer ( that would be the gray matter, I think ). Anyways, if you can get that book, it's quite interesting, as it proposes to add another cybernetic layer. About the other question, I think that the dificult thing is to measure the quality of the brain. You cannot say, as is the case with peas, that the brain is yellow, or big, or small. The only thing you can do is to give some benchmarks, like the several IQ. So, the first thing before measuring genetic inheritance in the brain is to establish a set of parameters to measure it. I don't know if this answer to any of your questions, but in any case it may also be a matter of discussion. JJ Merelo JMERELO@UGR.ES ------------------------------ Subject: how our genes give us brains From: Stephen Smoliar <smoliar@vaxa.isi.edu> Date: Thu, 12 Jul 90 11:38:35 -0700 Kingsley Morse's question of how our genes ultimately provide us with brains structured the way they are is one which has been pursued by Gerald Edelman. His book NEURAL DARWINISM makes a case for the argument that ALL physiological structure arises from a variety of selective processes. His concrete examples are concerned with chick feathers, but he extrapolates the argument to include not only the general architecture of the brain but also the neuronal wiring therein. This material is apparently discussed at greater length in his subsequent book, TOPOBIOLOGY; but I have not yet had a chance to look at that one. ------------------------------ Subject: Anybody in Poland working in NN From: JJ Merelo <jmerelo@ugr.es> Date: 11 Jul 90 16:07:00 +0200 I am going to Poland very soon on a tourist visit, and I would like to contact with anybody working on the same subject. If there is anybody in Cracow or Warszaw working on any aspect of Neural Networks, please e-mail to JJ Merelo JMERELO@UGR.ES ------------------------------ Subject: Cohen, Dunbar, & McClelland already published? From: Sven Blankenberger <I3160903%DBSTU1.BITNET@CUNYVM.CUNY.EDU> Organization: Dept. of Psycholoy, University of Braunschweig Date: Wed, 11 Jul 90 11:17:14 -0500 In their 1989 Psychological Review article Seidenberg & McClelland cited the following: Cohen, J., Dunbar, K., & McClelland, J. L. (1989) On the control of automatic processes: A parallel distributed processing model of the stroop task. Manuskript submitted for publication. Does anyone know where to find this article? If yes, would you please e-mail the complete reference, otherwise would you please e-mail the e-mail address of one of the authers. Many thanks Sven Sven Blankenberger (e-mail: i3160903@dbstu1.bitnet) Dept. of Psychology University of Braunschweig Spielmannstr. 19 3300 Braunschweig, West Germany ------------------------------ Subject: job opening From: Mike Cohen <mike@speech.sri.com> Date: Wed, 11 Jul 90 14:06:18 -0700 Their is a job opening on the research staff at SRI to participate in research in neural nets applied to computer speech recognition. The qualifications, in order of priority, include: Background in neural nets Strong C programming skills Background in speech recognition MS/PhD For information, contact: Dr. Michael H. Cohen SRI International, Rm EK182 333 Ravenswood Ave. Menlo Park, CA 94025 (415) 859-5977 mcohen@speech.sri.com ------------------------------ Subject: Evolving Networks - New TR From: rbelew@UCSD.EDU (Rik Belew) Date: Tue, 26 Jun 90 05:26:18 -0700 EVOLVING NETWORKS: USING THE GENETIC ALGORITHM WITH CONNECTIONIST LEARNING Richard K. Belew John McInerney Nicolaus Schraudolf Cognitive Computer Science Research Group Computer Science & Engr. Dept. (C-014) Univ. California at San Diego La Jolla, CA 92093 rik@cs.ucsd.edu CSE Technical Report #CS90-174 June, 1990 ABSTRACT It is appealing to consider hybrids of neural-network learning algorithms with evolutionary search procedures, simply because Nature has so successfully done so. In fact, computational models of learning and evolution offer theoretical biology new tools for addressing questions about Nature that have dogged that field since Darwin. The concern of this paper, however, is strictly artificial: Can hybrids of connectionist learning algorithms and genetic algorithms produce more efficient and effective algorithms than either technique applied in isolation? The paper begins with a survey of recent work (by us and others) that combines Holland's Genetic Algorithm (GA) with connectionist techniques and delineates some of the basic design problems these hybrids share. This analysis suggests the dangers of overly literal representations of the network on the genome (e.g., encoding each weight explicitly). A preliminary set of experiments that use the GA to find unusual but successful values for BP parameters (learning rate, momentum) are also reported. The focus of the report is a series of experiments that use the GA to explore the space of initial weight values, from which two different gradient techniques (conjugate gradient and back propagation) are then allowed to optimize. We find that use of the GA provides much greater confidence in the face of the stochastic variation that can plague gradient techniques, and can also allow training times to be reduced by as much as two orders of magnitude. Computational trade-offs between BP and the GA are considered, including discussion of a software facility that exploits the parallelism inherent in GA/BP hybrids. This evidence leads us to conclude that the GA's GLOBAL SAMPLING characteristics compliment connectionist LOCAL SEARCH techniques well, leading to efficient and reliable hybrids. -------------------------------------------------- If possible, please obtain a postscript version of this technical report from the pub/neuroprose directory at cheops.cis.ohio-state.edu. Here are the directions: /*** Note: This file is not yet in place. Give us a few days, ***/ /*** say until after 4th of July weekend, before you try to get it. ***/ unix> ftp cheops.cis.ohio-state.edu # (or ftp 128.146.8.62) Name (cheops.cis.ohio-state.edu:): anonymous Password (cheops.cis.ohio-state.edu:anonymous): neuron ftp> cd pub/neuroprose ftp> type binary ftp> get (remote-file) evol-net.ps.Z (local-file) foo.ps.Z ftp> quit unix> uncompress foo.ps.Z unix> lpr -P(your_local_postscript_printer) foo.ps /*** Note: This file is not yet in place. Give us a few days, ***/ /*** say until after 4th of July weekend, before you try to get it. ***/ If you do not have access to a postscript printer, copies of this technical report can be obtained by sending requests to: Kathleen Hutcheson CSE Department (C-014) Univ. Calif. -- San Diego La Jolla, CA 92093 Ask for CSE Technical Report #CS90-174, and enclose $3.00 to cover the cost of publication and postage. ------------------------------ Subject: Tech Reports Available From: Eduardo Sontag <sontag@hilbert.RUTGERS.EDU> Date: Wed, 27 Jun 90 16:27:35 -0400 The following report is now available: "On the recognition capabilities of feedforward nets" by Eduardo D. Sontag, SYCON Center, Rutgers University. ABSTRACT: In this note we deal with the recognition capabilities of various feedforward neural net architectures, analyzing the effect of direct input to output connections and comparing Heaviside (threshold) with sigmoidal response units. The results state, roughly, that allowing direct connections or allowing sigmoidal responses doubles the recognition power of the standard architecture (no connections, Heaviside responses) which is often assumed in theoretical studies. Recognition power is expressed in terms of various measures, including worst-case and VC-dimension, though in the latter case, only results for subsets of the plane are proved (the general case is still open). There is also some discussion of Boolean recognition problems, including the example of computing N-bit parity with about N/2 sigmoids. --------------------------------------------------------------------------- To obtain copies of the postscript file, please use Jordan Pollack's service: Example: unix> ftp cheops.cis.ohio-state.edu # (or ftp 128.146.8.62) Name (cheops.cis.ohio-state.edu:): anonymous Password (cheops.cis.ohio-state.edu:anonymous): <ret> ftp> cd pub/neuroprose ftp> binary ftp> get (remote-file) sontag.capabilities.ps.Z (local-file) foo.ps.Z ftp> quit unix> uncompress foo.ps unix> lpr -P(your_local_postscript_printer) foo.ps ---------------------------------------------------------------------------- If you have any difficulties with the above, please send e-mail to sontag@hilbert.rutgers.edu. DO NOT "reply" to this message, please. ------------------------------ Subject: Report on Non Linear Optimization From: Manoel Fernando Tenorio <tenorio@ee.ecn.purdue.edu> Date: Mon, 02 Jul 90 11:06:05 -0500 This report is now available from Purdue University. There is a fee for overseas hardcopies. An electronic copy will be soon available at the Ohio database for ftp. Please send your request to Jerry Dixon (jld@ee.ecn.purdue.edu). Do not reply to this message. COMPUTATIONAL PROPERTIES OF GENERALIZED HOPFIELD NETWORKS APPLIED TO NONLINEAR OPTIMIZATION Anthanasios G. Tsirukis and Gintaras V. Reklaitis School of Chemical Engineering Manoel F. Tenorio School of Electrical Engineering Technical Report TREE 89-69 Parallel Distributed Structures Laboratory School of Electrical Engineering Purdue University ABSTRACT A nonlinear neural framework, called the Generalized Hopfield Network, is proposed, which is able to solve in a parallel distributed manner systems of nonlinear equations. The method is applied to the general optimization problem. We demonstrate GHNs implementing the three most important optimization algorithms, named the Augmented Lagrangian, Generalized Reduced Gradient and Successive Quadratic Programming methods. The study results in a dynamic view of the optimization problem and offers a straightforward model for the parallelization of the optimization computations, thus significantly extending the practical limits of problems that can be formulated as an optimization problem and which can gain from the introduction of nonlinearities in their structure (eg. pattern recognition, supervised learning, design of content-addressable memories). ------------------------------ Subject: Call for Participation in Connectionist Natural Language Processing From: cpd@aic.hrl.hac.com Date: Tue, 10 Jul 90 10:52:42 -0700 AAAI Spring Symposium Connectionist Natural Language Processing Recent results have lead some researchers to propose that connectionism is an alternative to AI/Linguistic approaches to natural language processing, both as a cognitive model and for practical applications. This symposium will bring together both critics and proponents of connectionist NLP to discuss its strengths and weaknesses. This symposium will cover a number of areas, spanning from new phonology models to connectionist treatments of anaphora and discourse issues. Participants should address what is new that connectionism brings to the study of language. The purpose of the symposium is to examine this issue from a range of perspectives including: Spoken language understanding/generation Parsing Semantics Pragmatics Language acquisition Linguistic and representational capacity issues Applications Some of the questions expecting to be addressed include: What mechanisms/representations from AI/Linguistics are necessary for connectionist NLP? Why? Can connectionism help integrate signal processing with knowledge of language? What does connectionism add to other theories of semantics? Do connectionist theories have implications for psycholinguistics? Prospective participants are encouraged to contact a member of the program committee to obtain a more detailed description of the symposium's goals and issues. Those interested in participating in this symposium are asked to submit a 1-2 page position paper abstract and a list of relevant publications. Abstracts of work in progress are encouraged, and potential participants may also include 3 copies of a full length paper describing previous work. Submitted papers or abstracts will be included in the symposium working notes, and participants will be asked to participate in panel discussions. Three (3) copies of each submission should be sent to arrive by November 16, 1990 to: Charles Dolan, Hughes Research Laboratories, RL96, 3011 Malibu Canyon Road, Malibu CA, 90265 All submissions will be promptly acknowledged. E-Mail inquiries may be sent to: cpd@aic.hrl.hac.com Program Committee: Robert Allen, Charles Dolan (chair), James McClelland, Peter Norvig, and Jordan Pollack. ------------------------------ Subject: Apologies on Call for Paticipation From: cpd@aic.hrl.hac.com Date: Tue, 10 Jul 90 16:52:56 -0700 My previous message left out some information. AAAI Spring Symposium March 26-28 Stanford University Connectionist Natural Language Processing ... from previous message... All submissions will be promptly acknowledged and registration materials will be sent to authors who submit abstracts. The Spring Symposium Series is an annual event of AAAI. Members of AAAI will receive a mailing from them. Attendance is by submission and acceptance of materials only, except that AAAI will fill available spaces if the program committee does not select enough people. Enough people is 30-40. No proceedings will not be published, but working papers, with submissions from attendees, will be distributed at the Symposium. Sorry for the momentary confusion -Charlie Dolan ------------------------------ Subject: Technology Transfer Mailing List From: Bill Hefley <weh@SEI.CMU.EDU> Date: Tue, 10 Jul 90 21:13:39 -0400 The Technology Applications group of the Software Engineering Institute is pleased to announce the creation of a new electronic mailing list: technology-transfer-list. This mailing list, focused on technology transfer and related topics, is intended to foster discussion among researchers and practitioners from government and industry who are working on technology transfer and innovation. Relevant topics include: -- organizational issues (structural and behavioral) -- techno-economic issues -- business and legal issues, such as patents, licensing, copyright, and commercialization -- technology transfer policy -- technology maturation to support technology transition -- lessons learned -- domestic and international technology transfer -- transition of technology from R&D to practice -- planning for technology transition -- models of technology transfer -- studies regarding any of these topics The technology-transfer-list is currently not moderated, but may be moderated or digested in the future if the volume of submissions warrants. The electronic mail address for submissions is: technology-transfer-list@sei.cmu.edu To request to be added to or dropped from the list, please send mail to: technology-transfer-list-request@sei.cmu.edu Please include the words "ADD" or "REMOVE" in your subject line. Other administrative matters or questions should also be addressed to: technology-transfer-list-request@sei.cmu.edu The SEI is pleased to provide the facilities to make this mailing list possible. The technology-transfer-list is the result of two SEI activities: -- transitioning technology to improve the general practice of software engineering -- collaborating with the Computer Resource Management Technology program of the U.S. Air Force to transition technology into Air Force practice The SEI is a federally funded research and development center sponsored by the U.S. Department of Defense under contract to Carnegie Mellon University. ------------------------------ End of Neuron Digest [Volume 6 Issue 43] ****************************************