neuron-request@HPLABS.HP.COM ("Neuron-Digest Moderator Peter Marvit") (07/14/89)
Neuron Digest Thursday, 13 Jul 1989 Volume 5 : Issue 30 Today's Topics: Gradient descent updates Pointer to Bill Polkingham References wanted for non-iterative training Request for ART2 info request for literature on visualization of neural nets Special Interest Group Meetings in Winter IJCNN 90 Conference Spin Glass and Neural Networks Re: Spin Glass and Neural Networks Re: Spin Glass and Neural Networks Re: Spin Glass and Neural Networks Submission - List of Neural Network Methods Summary: Help: Neural Nets/Cell-Automata/Dynamic Systems ... Re: Help: Neural Nets/Cell-Automata/Dynamic Systems ... RE: Help: Neural Nets/Cell-Automata/Dynamic Systems ... Send submissions, questions, address maintenance and requests for old issues to "neuron-request@hplabs.hp.com" or "{any backbone,uunet}!hplabs!neuron-request" ARPANET users can get old issues via ftp from hplpm.hpl.hp.com (15.255.16.205). ------------------------------------------------------------ Subject: Gradient descent updates From: leary@Luac.Sdsc.Edu (Bob Leary) Date: Fri, 07 Jul 89 23:03:54 +0000 With regard to Francesco Camargo's recent question concerning whether in gradient descent procedures such as back-propagation it is better to cycle through all input-output pairs before making updates, or to update after each pair: There is a special case where precise analytical results are known. Consider the solution of n linear equations in n unknowns where the coefficient matrix is diagonally dominant - i.e. the coefficient of the ith variable in the ith equation is much larger in magnitude than the other coefficients in that equation (how much is "much" can be found in any numerical analysis text). Two so-called iterative methods called Jacobi and Gauss-Seidel are essentially gradient descent techniques that minimize the sum of the squares of the errors contributed by each equation, where the error is the difference between the desired right hand side and what you actually get by plugging the current approximate solution into the left hand side. Each equation is presented in turn in a cyclic pattern. When the ith equation is presented, an adjustment is made in the ith variable so as to make the error for that equation equal to zero (of course, this adjustment also affects the errors contributed by all the other equations). The only difference between the two methods is that with Jacobi, the errors are only updated after a complete cycle, while with Gauss-Seidel, the updating is done after each presentation. The upshot of all this is that both methods converge under exactly the same conditions, and that the convergence rates are known (again, consult that numerical analysis text) - the Gauss-Seidel method wins by a landslide. Bob Leary San Diego Supercomputer Center leary@sds.sdsc.edu ------------------------------ Subject: Pointer to Bill Polkingham From: cditi!sdp@uunet.UU.NET (Steve Poling) Date: Tue, 11 Jul 89 10:57:53 -0400 Peter, Many thanx for your neural networks digest. You refered to a quarterly service where recently awarded patents about neural networks are forwarded to subscribers. The service is offered by a Bill Polkinghorn son of the fellow who was running the INNS sign-up desk. The company is called "MIPPS" in capital greek letters. The address of this service is: Bill Polkinghorn P.O. Box 15226 Arlington, VA 22215 Aside to a similarity in last names (Poling and Polkinghorn) I have no connection or commercial interest in this company. Cheers, Steve Poling CDI Technologies ------------------------------ Subject: References wanted for non-iterative training From: Kemp@DOCKMASTER.NCSC.MIL Date: Fri, 07 Jul 89 20:33:00 -0400 A recent trade publication (EE Times, June 26, p.36) had an article about a commercial NN product that trains "1000 times quicker" than iterative methods, using techniques from applied mathematics to solve for weights directly. From the article (without permission): "The first pass throught the data builds a giant matrix mapping all of the inputs. The second pass performs a non-linear inversion of that matrix, resulting in the weight matrix for the neural network." . . . "Back propogation networks are only a little bit non-linear, and we were able to find several non-linear variants of matrix inversion that do the job for most cases." Is there any published work describing applications of non-linear algebra to NN training, and the types of problems to which it might be suited. Dave Kemp <Kemp@dockmaster.ncsc.mil> ------------------------------ Subject: Request for ART2 info From: Jeff Kowing <kowing%nasa-jsc.csnet@RELAY.CS.NET> Date: Mon, 12 Jun 89 16:55:28 -0500 I am looking for literature analyzing, either mathematically or emperically, the performance of Grossberg's ART2 paradigm. If anyone is aware of such literature please send references to kowing@nasa.jsc.gov Thanks alot, I appreciate your time and help! Jeff Kowing NASA/Johnson Space Center P.S. I already have the ART2 article from the Dec. 1987 issue of Applied Optics. ------------------------------ Subject: request for literature on visualization of neural nets From: mcvax!ethz!mr@uunet.uu.net Date: 28 Jun 89 18:22:00 -0800 I'm looking for any work done in the field of visualization of neural nets, ie. any help to intuitively see what a neural net is doing, how it learns, what features it detects etc. Please e-mail to me; I'll summarize if there is enough interest. Thanks in advance. Marc Raths, Swiss Federal Institute of Technology UUCP/EUNET: mr@ethz.uucp or Rigistrasse 53 ...!{uunet,seismo,ukc}!mcvax!cernvax!ethz!mr CH-8006 Zurich, Switzerland CSNET/ARPA : mr%ethz.uucp%ifi.ethz.ch@RELAY.CS.NET Voice : +41-1-361 5575 BITNET/EARN: mr%ethz.uucp@cernvax.BITNET ------------------------------ Subject: Special Interest Group Meetings in Winter IJCNN 90 Conference From: fong@sun.soe.clarkson.edu Organization: Clarkson University, Potsdam, NY Date: Tue, 27 Jun 89 13:41:13 +0000 In the upcoming International Joint Neural Network Conferenece in January in Washington DC, there will be a time set aside for Special Interest Group meetings. (During one of the dinner time or lunch time). The purpose of this posting is to call the attention of all the SIG points of contact who have formed a SIG during or after the Boston ISNN Annual meeting to consider this meeting, and desseminate the information to your members. 1. Please submit a one-page proposal of activity for your SIG meeting to Harold Szu, Local Organizing Chair, IJCNN 90. His FAX numbers are: 202-767-4277 or 202-767-1494. 2. You can contact Maureen Caudill at 619-485-1809 for room arrangement for your group meeting. If you have difficulty contacting Maureen, relay the information to Harold Szu. 3. Please remind your group member and colleagues that during this meeting, several Nobel Lureates are invited to give talks, and there are several technical tracks (application, biology and theory) for paper submission and presentation. Please plan to participate this exciting conference. David Yushan Fong ------------------------------ Subject: Spin Glass and Neural Networks From: hitomi@ogccse.ogc.edu (Hitomi Ohkawa) Organization: Oregon Graduate Center, Beaverton, OR Date: Fri, 09 Jun 89 19:26:33 +0000 I am interested in finding out if there are any on-going research efforts that take a field-theoretic approach to studying behaviors of neural networks. I recently had a chance to look through a book titled "Spin Glasses and Other Frustrated Systems" by D. Chowdhury, and intrigued by a wide applicability of a SG system, from a travelling salesman problem to neural networks. I used to study physics, and once tried to apply the Ising model to a certain physical system. Now being in the field of computer science (though not a neural network specialist), it is very interesting to see a concept from physics familiar to me being applied to a certain form of computer architecture. Any information regarding this particular area (research efforts, references, good textbooks on the subject, etc.) is greatly appreciated. Below is my address. Hitomi Ohkawa Dept. of Computer Science and Engineering Oregon Graduate Center 19600 N.W. Von Neumann Drive Beaverton, OR. 97006-1999 (503) 690-1151 hitomi@cse.ogc.edu (CSNET) Thank you very much in advance. ------------------------------ Subject: Re: Spin Glass and Neural Networks From: giant@lindy.Stanford.EDU (Buc Richards) Organization: Stanford University Date: Fri, 09 Jun 89 20:51:30 +0000 The analogy to Spin Glasses is used in the Stochastic neural network of Boltzmann machines. I do not how to give a quick explanation, but this is discussed in Emile Aarts and Jan Kort's excellent new book, "Simulated Annealing and Boltzmann Machines" starting on page 148. However, depending on your previous knowledge of Simulated Annealing, Boltzmann Machines, and Spin Glasses it may take more than a few pages of reading to appreciate the analogy. The book has just been published by John Wiley & Sons (New York), so it may not yet be available at your library. Rob Richards Stanford University ------------------------------ Subject: Re: Spin Glass and Neural Networks From: bph@buengc.BU.EDU (Blair P. Houghton) Organization: Boston Univ. Col. of Eng. Date: Sat, 10 Jun 89 18:04:11 +0000 In addition, you could investigate the articles by Amit, Gutfreund, and Sompolinski, ca. 1984 (I think, could be anywhere 1982-1987) which appeared (I can't remember where. Could have been IEEE Systems, Man, and Cybernetics, or it could have been a physics journal. The Science Citation Index would defininitely have them, since that's where I originally found the reference.) There were a few of them, and were inspired by Hopfield's paper on the TSP problem (which means it couldn't be 1982...) --Blair "...I have it here _somewhere_..." ------------------------------ Subject: Re: Spin Glass and Neural Networks From: wine@maui.cs.ucla.edu (David Wine) Organization: UCLA Computer Science Department Date: Wed, 14 Jun 89 00:34:00 +0000 reference: Haim Sompolinsky, Statistical Mechanics of Neural Networks, Physics Today, December, 1988 ------------------------------ Subject: Submission - List of Neural Network Methods From: David Kanecki <kanecki@vacs.uwp.wisc.edu> Date: Fri, 07 Jul 89 22:20:04 -0500 Dear Peter, Enclosed is article summarizing the neural network methods in use. This list was composed from articles I received and responses received from e-mail. Also, I found the articles of neural networks in chemistry from Chemical and Engineering New and Proceeding of the National Academy of science the most informative as to method, set up, and results. I think its great the neuron digest can be the sentinel, source, and sound post for new ideas and concepts. Keep up the good work. David Kanecki kanecki@vacs.uwp.wisc.edu P.O. Box 93 Kenosha, WI 53141 Article Enclosure: From information received from articles and responses for information from various people I have compiled a list of neural network methods currently in use. The methods in use are: 1. Generalized Delta Rule, digital neuron B(I)= A(J)*W(I,J) 2. Delta with transfer function 1, digital and analog neuron B2(I) = A(J)*W(I,J) B(I) = 1/(1+exp(-B2(I)) 3. Skeletonization, digital and analog neuron The neuron that does not change the state of the output neuron is turned off in the update procedure. Thus, only neurons that affect the state of the output neurons are included in the error E(I) used in the update procedure to produce a new matrix W(I,J). 4. Genetic methods, 9 classes, digital and analog Based on test using 1 of the 9 methods the genetic method modifies the value in the error vector using a specific criteria. This is analogous to the proof reading done by DNA polyomerase in the cell. Based on experiments I have performed using 1 of the 9 genetic methods, the genetic method causes a noisy /conflicting network to converge faster than compared to using the generalized delta rule. But, on a network where the noise or conflict is minimal the genetic method is slower than the delta rule on the rate of convergence. I considered a network to have the minimal noise or conflict if the error rate after 50 trails was less than 0.8 percent. 5. Winner take all, digital and analog The output neuron with the biggest output is allowed to fire and be updated. The other neurons are turned off. This method has been used by a group of researchers to model biological neural activity in the olfactory region of the brain. 6. Back propogation, Digital and analog This method uses the delta rule but updates the W(I,J) matrix by using the primary error vector E(I) and the first derivative dW(I,J) of the neural matrix. If you would like to discuss these methods or have comments please contact me by e-mail or write. I offer custom analysis services and programming. If one is interested in this type of work send me the details and overview for a free estimate. Currently, I am a recent graduate of the University of Wisconsin with degrees in Applied Computer Science and Biological Science. I am looking for work in computer science, mathematics, biology, statistics, artificial intelligence, or neural networks. Lastly, I would be interested in making a survey of which neural network methods are being used. And, I would provide a tabulation of the results as they are received. I can be contacted at the address below: David H. Kanecki P.O. Box 93 Kenosha, WI 53141 Bitnet: kanecki@vacs.uwp.wisc.edu P.S. I have developed a analog and digital neural network programmer and simulator available for mainframes, mini, and micro computers [[...]]. For data sheets, product information [[ ,PRICE LIST, AND CREDIT INFORMATION ]] please write to the address above. [[ Editor's Note: This Digest, in keeping with ARPANET guidelines, attempts to be as non-commercial as possible while still providing information which will benefit the greatest number of people. As I've mentioned before, I exercise some limited editing on "commericials," and have trimmed some of the text above accordingly. Please contact David for more information about his product and services. If anyone has comments on my (admittedly arbitrary) editing policy, eitehr pro or con, please send e-mail. -PM ]] ------------------------------ Subject: Summary: Help: Neural Nets/Cell-Automata/Dynamic Systems ... From: Darrell Schiebel <unmvax!aplcen!haven!h.cs.wvu.wvnet.edu!cerc.wvu.wvnet.edu!drs@UCBVAX.BERKELEY.EDU> Date: 06 Jul 89 16:19:58 +0000 Some time ago I posted a message asking for help: >I am interested in the Following: > > Computing with: > Neural Networks > Cellular Automata > Dynamic Systems > Self Organizing Systems > etc ... I received several responses, and several people ask for a posting of a summary. I would like to thank the people who responded with the insightful information which follows; It was a great help to me, and perhaps, it will aid others. Tim Swenson (tswenson@daitc.mil) Charlie Sorsby (sorsby@hi.unm.edu) Tony Meadors (meadors@cogsci.ucsd.edu) Hal Hardenbergh (236) (hal@vicom.com) Cliff Joslyn (vu0112@bingvaxu.cc.binghamton.edu) Dave Hiebeler (hiebeler@cs.rpi.edu) Sue Worden (worden@ut-emx.UUCP) Russell Beale (russell@minster.york.ac.uk) - -------------------------------------------------------------------- The EECE Dept. at the Univeresity of New Mexico is doing some work in the NN area. The two professors who are doing NN stuff in the Dept. are Dr. Don Hush and Dr. Victor Bolie. - -------------------------------------------------------------------- These areas, are quite broad and differ widely in character depending on the particular projects they are applied. What I mean is that the relevance of particular books, classes, or mathematical methods is almost entirely dependent on what you intend to UNDERSTAND or ENGINEER. The practical reason for including your reseach intention so early in the process is this; if you go to someone in say the applied engineering department and ask how you should learn about control systems they will gladly point the direction, the physics department another, the psycholgy department differently yet, and so on. And even within those disciplines, say psychology, control system principles (just my example) underlie models which otherwise have little or no relation to one another (say the study of reaching for moving targets vs. models of motivation). The bottom line...go as directly as possible to the literature relevant to what you wish to accomplish and from that you will learn what background and related topics you need to master. - -------------------------------------------------------------------- This is the best available practical book on neural nets: "Neural Computing: Theory and Practice," Philip D. Wasserman Van Nostrand-Reinhold 1989 This slender paperback contains Richard Lippmann' tutorial (Apr 1987 IEEE ASSP) reprint, and is recommended for that reason: "Artificial Neural Networks: Theoretical Concepts," V. Vemuri editor (IEEE) Computer Society Press, 1988 Computer Society order #855; IEEE Catalog # EH 0279-0 The next book is a huge $55 volume, and is utterly invaluable if you are interested in the historical background of anns. If you aren't interested in the historical background, don't buy it. "Neurocomputing: Foundations of Research," Anderson and Rosenfeld, editors MIT Press, 1988 Once you decide you are serious about anns, these two collections of technical papers will collectively set you back about $60, and are worth it: "Proceedings of the 1988 Connectionist Models Summer School (Carnegie-Mellon)" Morgan Kaufmann Publishers 1989 Paperback "Advances in Neural Information Processing Systems, Vol 1," D.S. Touretzky ed Morgan Kaufmann Publishers 1989 Hardcover This is a $10 paperback which is a reprint of the (1988?) MIT house organ "DAEDALUS" magazine special issue on AI. The leadoff article by Papert is hilarious if you are a backprop fan. Many, and diverse, opinions, some of them frankly hostile to AI: "The Artificial Intelligence Debate," Stephen R. Graubard editor MIT Press 1988 There is the hardware and there is the wetware. The dividing line between anns and cognitive sciences is not well defined. The papers in this book lean in the direction of wetware. "Neural Networks and Natural Intelligence," Stephen Grossberg editor - -------------------------------------------------------------------- Two good books about cellular automata: Cellular Automata Machines: A New Environment for Modeling by T. Toffoli and N. Margolus, MIT Press, 1986 (or maybe 1987) -- good intro to CA applied to physical modeling, and CA in general Theory and Applications of Cellular Automata edited by S. Wolfram, Scientific Press, 1984 (not sure of publisher/date) -- a collection of articles (many by Wolfram) about just what the title says. Not as light reading as the first book. - -------------------------------------------------------------------- I hardly ever see the following book referenced, but I think it might provide a reasonable introduction to some of your areas of interest: Glorioso, Robert M. and Fernando C. Colon Osorio ENGINEERING INTELLIGENT SYSTEMS : Concepts, Theory, and Applications Digital Equipment Corporation; Bedford, Massachusetts; 1980 ISBN 0-932376-06-1; 472 pages Abbreviated Table of Contents Chap 1 : Computers and Intelligence Chap 2 : Game Playing and Machines Chap 3 : Reason, Logic, and Mathematics Chap 4 : Computers and Automata Chap 5 : Adaption, Learning, Self-Repair, and Self-Organization Chap 6 : Stochastic Automata Models Chap 7 : Adaptive, Learning, and Self-Organizing Controllers Chap 8 : Cybernetic Techniques in Communication Systems Chap 9 : Stochastic Automata Models in Computer and Communication Networks Chap 10 : Reliability and Repair Chap 11 : Neurons and Neural Models Chap 12 : Threshold Logic Chap 13 : Pattern Recognition Chap 14 : Computer Vision Chap 15 : Robotics From your posting, I gather that your orientation is toward a blend of computer science, computer engineering, and linear/non-linear systems theory and engineering. That in itself indicates that you are probably seeking a university with faculty/student/program crossovers between appropriate academic departments. The University of Texas at Austin is one such university. For details, write: Dean of Graduate Studies Main Building 101 The University of Texas at Austin Austin, Texas 78712 Finally, in whatever graduate program you finally choose, I encourage you to set aside a few course hours for psychology (cognitive science), neuroanatomy/neurophysiology, et cetera. The organic perspective gained on our technological pursuits is invaluable. - -------------------------------------------------------------------- Parallel Distributed Processing: D. E. Rumelhart, J. L. McClelland Explorations in the Microstructure of Cognition MIT Press, Cambridge Mass., 1986. 3 vols, excellent. An Introduction to Computing with Neural Nets Richard P. Lippmann IEEE ASSP Magazine 4-22 April 1987 Hopfield hamming carpenter grossberg perceptron organizing kohonen introduction Good clear introduction, with many references. An Introduction to Neural Computing Teuvo Kohonen Neural Networks V 1 N 1 P 3-16 D 1988 Perceptrons M Minsky, S. Papert 1969 MIT Press Contains criticisnm of single layered networks. - -------------------------------------------------------------------- Sci American Either Computer Rec or Math column. August 1988, May 1985, Mar 1984, Feb 1971, Oct. 1970 Wolfram, Stephen Los Alamos Science fall 1983 "Cellular Automata" Cooper, Necia Los Alamos Science Fall 1983, "From Turing and Von Neuman to the Present" Buckingham, David, "some facts on life", Byte Dec. 1979 The Wolfram article is very good. ------------------------------ End of Neurons Digest *********************