neuron-request@HPLABS.HP.COM (Neuron-Digest Moderator Peter Marvit) (11/09/88)
Neuron Digest Tuesday, 8 Nov 1988 Volume 4 : Issue 19 Today's Topics: Administrivia Re: ART source readings? Re: ART source readings? Book Announcement Re: Cyberspace Implementation Issues Learning with NNs Re: MacBrain Neural Network Companies Re: Neural Network Companies Re: Neural Network Companies Neuron resolution NIPS computer demos Outlets for theoretical work Students for a Better NN Class Suggestions needed... Re: Wanted: info about GENESIS program 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 From: Neuron-Digest Moderator Peter Marvit Date: Tue, 8 Nov 88 11:31:27 pst [[ The digest has been a bit delayed, partially due to the "worm" which caused our site to be isolted from the Internet until we were sure that it was under control. I'll try to get up to date in the next week. In the mean time, any mail you might have sent could have bounced. Please send it again. I received one response to my "editorial" from last issue. I welcome additional feedback as to editorial policy. -PM ]] ------------------------------------------------------------ Subject: Re: ART source readings? From: demers@beowulf.ucsd.edu (David E Demers) Organization: EE/CS Dept. U.C. San Diego Date: 23 Oct 88 00:58:04 +0000 >Could someone please email (or otherwise send) me references to the >basic papers, books, whatever about Grossberg's ART? I thought I'd post my response because it may be of wide interest. Perhaps the best place to start on ART would be a couple of papers from ICNN-87, found in the proceedings: Carpenter & Grossberg, ART 2: Self-Organization of stable category recognition codes for analog input patterns. Carpenter & Grossberg, Invariant pattern recognition and recall by an attentive self-organizing ART architecture in a nonstationary world. as well as some other papers from the same session. For the beginnings, I think Grossberg, Adaptive pattern classification and universal recoding, II: Feedback, expectation, olfaction and illusions. in Biological Cybernetics, 23, (1976) 187-202. is a good paper showing the underpinnings of ART. I think that this paper is reprinted in the collection of seminal works put together by James Anderson, Neurocomputing: Foundations of Research. MIT Press, 1988. The bibliography at the end of the first two papers, above, will give you some more to look at. *Personal caveat* I think Grossberg's work is very important, however, it takes a long time to read his papers. They are pithy... Dave DeMers UCSD Dept. of Computer Science & Engineering La Jolla, CA 92093 (619) 534-6254 demers@cs.ucsd.edu ------------------------------ Subject: Re: ART source readings? From: bph@buengc.BU.EDU (Blair P. Houghton) Organization: Boston Univ. Col. of Eng. Date: 26 Oct 88 01:47:29 +0000 >>Could someone please email (or otherwise send) me references to the >>basic papers, books, whatever about Grossberg's ART? Wot Luck! I was flipping through Kandel and Schwarz just now, and what should fall out but a library Recall Notice for Grossberg's compendium of papers, _Neural_Networks_and_Natural_Intelligence_, call number QP 363.3 N44 --Blair "Don't recall me, I'll recall you, if I can get my Natural Intelligence to adapt and resonate simultaneously..." ------------------------------ Subject: Book Announcement From: pattis@june.cs.washington.edu (Richard Pattis) Organization: U of Washington, Computer Science, Seattle Date: 26 Oct 88 04:20:38 +0000 A new book, called "Cognizers: Neural Networks and Machines that Think" about neural networks and the community that uses them has just become available at my local bookstore. The authors are Johnson and Brown. The publisher is John Wiley and Sons. The title should tell you where the authors are going. ------------------------------ Subject: Re: Cyberspace Implementation Issues From: jdb9608@ultb.UUCP (J.D. Beutel ) Organization: South Henrietta Institute of Technology (Info Systems) Date: 16 Oct 88 20:17:55 +0000 In article <10044@srcsip.UUCP> lowry@srcsip.UUCP () writes: > > >There's been a lot of discussion recently on how something (kind of) like >c-space might be implemented. The conventional wisdom seems to be that >you'd need super-high res color monitors and a graphics supercomputer >providing real-time images. > >It seems to me that kind of equipment would only be needed if you were >going to funnel all the info into an eye. I recall reading somewhere >that the nerves behind the retina do preprocessing on images before >sending the data down the optic nerve. If you could "tee" into the >optic nerve, it seems like you could feed in pre-digested data at >a much lower rate. > >Apologies if this idea/topic has been previously beaten to death. Beaten to death? Nonsense! I've heard alot about neural networks, artificial retinas in particular. Research is producing, on the dry side, theories about how machines can see, and conversly, on the wet side, how we ourselves see. All the theories I've heard of concure that the neurons which react immediately to light are input as groups to other neurons which react to higher forms like lines and dots and movement. But, while I think that the resultant information is more useful, I'd also guess that there is more of that information than there was raw information from which it was derived. For example: the silicon retina that some company (I don't remember the name) is working on with Carver Mead: every 6 light-sensitive neurons are sampled by 3 edge-sensitive neurons (up:down, (PI/4):(5PI/4), and (3PI/4):(7PI/4)). However, all the light-sensitive neurons are arranged in a hexigonal tessilate such that each neuron is part of 3 hexigons. Therefore, as the number of light-sensitive neurons increases, the ratio of edge-sensitive to light-sensitive approaches 1. Additionally, there are other higher forms, like dots and spots and motion in various directions, that will all be using those same light-sensitive neurons as input. That's why I think that "pre-digested data" might be 10 times more massive than the raw visual input. Of course, one could try to digest the data further, transmitting boxes and circles and motion paths as gestalt instead of transmitting the lines and corners that make them up. But, the further you digest the data, the deeper into the brain you must go. Pixels make up lines; lines make up corners; lines and corners make up squares and triangles; squares and triangles make up the picture of a house. The theories I've heard of agree that we are all born with the neurons pre-wired (or nearly so) to recognize lines, but I've heard of none that suggest that we are pre-wired with neurons that recognize a box with a triangle on top. Instead, we've learned to recognize a "house" because we've seen alot of them when we were young. The problem is that the way I learn "house" might be different from the way you learn "house." So, a video screen is a video screen to two different people, but a "tee into the optic nerve" would have to be very different for two different people, depending on how far back into the brain you jacked in. The system would have to be dynamic, since people learn as they age; what a house is to you at age 10 is not what a house is to you at age 20. Symstim and consensual hallucinations are taken for granted in cyberpunk, and I took them for granted too. The more I think about it, however, the less probable is seems. I'm cross-posting my lame followup to the neural-nets group in the hope that someone there will have some comment or idea on how a computer could possibly generate a consensual hallucination for its operator, hopefully entirely within the operator's mind, as opposed to controling holograms and force fields which the operator would 'really' see around him. 11011011 ------------------------------ Subject: Learning with NNs From: Dario Ringach <dario%TECHUNIX.BITNET@CUNYVM.CUNY.EDU> Date: Wed, 19 Oct 88 13:36:32 +0200 Has anyone tried to approach the problem of learning in NNs from a computability-theory point of view? For instance, let's suppose we use a multilayer perceptron for classification purposes. What is the class of discrimination functions learnable with a polynomial number of examples such that the probability of misclassification will be less than P (using a determined learning algorithm, such as back-prop)? It seems to me that these type of questions are of importance if we really want to compare between different learning algorithms, and computational models. Does anyone have references to such a work? Any references will be appreciated! Thanks in advance! Dario. - ------------------------------------------------------------- BITNET: dario@techunix | "Living backwards!" Alice repeated in great Dario Ringach | astonishment. "I never heard of such a thing!" Retner 12/7 | 32819 Haifa, | "--But there's one great adventage in it, that one's ISRAEL | memory works both ways" The Queen remarked. - ------------------------------------------------------------- ------------------------------ Subject: Re: MacBrain From: hsg@romeo.cs.duke.edu (Henry Greenside) Date: 25 Oct 88 19:48:31 +0000 I would recommend against anyone purchasing MacBrain. Several copies were purchased here at Duke University for evaluation. MacBrain had a reasonable mouse interface for placing neurons and for linking neurons, and had built-in rules for learning such as the delta-rule, back-propagation, and the Boltzmann machine. There was no language that allowed complicated networks to be set up, so drawing and linking more than about ten neurons was tedious and impractical. The program is slow, making scaling studies also impractical. The worst part of the product was that it was extremely buggy and many advertised features had not been implemented. Repeated calls to the makers of MacBrain at Neuronics led to promises that bug-free versions would be available any day now, but they never arrived. Neuronics also refused to refund our money. Try another product from more trustworthy company. Henry Greenside ------------------------------ Subject: Neural Network Companies From: rravula@wright.EDU (R. Ravula) Organization: Wright State University, Dayton OH, 45435 Date: 25 Oct 88 15:30:59 +0000 About a month ago, I asked for a list of neural network companies. I am posting the only response I got. - ------------------------------------------------------ From: ames!ucsd!pnet12.cts.com!bstev (Barry Stevens) To: pnet101!pnet01!crash!osu-cis!wright!rravula Subject: neural net companies Status: RO HHC, Inc. San Diego CA 619-546-8877 (Dr Robert Hecht-Nielsen) hardware - coprocessor board, supporting software SAIC Science Applications, Inc. San Diego - don't have phone hardware - coprocessor, supporting software Nestor, Inc. Providence, Rhode Island - don't have phone software - neural nets and applicationg AI Ware, Cleveland, Ohio - don't have phone coprocessor board, software NeuralTech, Portola Valley, CA - don't have phone software - "English" language for generating nets Applied AI Systems, Inc. (My company) San Diego consulting on commercial applications of neural nets, using them in a company. UUCP: {crash ncr-sd}!pnet12!bstev ARPA: crash!pnet12!bstev@nosc.mil INET: bstev@pnet12.cts.com - ------------------------------------------------------ Thank you, Mr. Stevens. - -- - -------------------------- Ramesh Ravula ------------------------ rravula%wright.edu@csnet-relay | Wright State Research Center ...!osu-cis!wright!rravula | 3171 Research Boulevard (513) 259-1392 | Kettering, OH 45420 ------------------------------ Subject: Re: Neural Network Companies From: rcsmith@anagld.UUCP (Ray Smith) Organization: Analytics, Inc., Columbia, MD Date: 27 Oct 88 14:27:23 +0000 In article <360@thor.wright.EDU> rravula@wright.EDU (R. Ravula) writes: >SAIC Science Applications, Inc. San Diego - don't have phone > hardware - coprocessor, supporting software Some info we have on SAIC follows: Science Applications International Corporation (SAIC) Sigma Neurocomputer Systems Division Jennifer Humphrey, Sales Manager 10260 Campus Point Drive (MS 71) San Diego, CA 92121 Voice: 619-546-6290 FAX: 619-546-6777 Hope this helps. Ray - -- =-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= Ray Smith | ...!uunet!mimsy!aplcen!\ Analytics, Inc. | ...!netsys!---anagld!rcsmith Suite 200 | ...!ethos! / 9891 Broken Land Parkway | Columbia, MD 21046 | Voice: (301) 381-4300 Fax: (301) 381-5173 =-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= ------------------------------ Subject: Re: Neural Network Companies From: muscarel@uicbert.eecs.uic.edu Date: 04 Nov 88 05:45:00 +0000 > /* ---------- "Neural Network Companies" ---------- */ > About a month ago, I asked for a list of neural network companies. I am > posting the only response I got. The magazine AI Expert has been running an ongoing series of articles on neural networks and connectionism over the past year or so. The August, 1988 issue (Vol 3, No 8) was totally devoted to this topic. The "Software Review" article in that issue (pp 73-85) contained a review entitled "12-Product Wrap-Up: Neural Networks" which contained descriptions and reviews of a number of neural network software products. Although the title indicates that 12 products are reviewed, an extensive table listing comparative features of the reviewed products has 13 entries. There is also a table of software vendors listing addresses and phone numbers. If you can't find this article I could mail you a copy. Tom Muscarello Dept EECS Electronic Mind Control Laboratory Univ. of IL at Chicago (mc/154) Chicago, IL 60680 : muscarel@uicbert.eecs.uic.edu ------------------------------ Subject: Neuron resolution From: rao@enuxha.eas.asu.edu (Arun Rao) Organization: Arizona State Univ, Tempe Date: 31 Oct 88 16:27:01 +0000 I'm making some studies on the theoretical capabilities of neural systems. I need information concerning the resolution of neurons. For example, what is the order of variance in measured firing frequencies ? Thiscould be a measure of inherent uncertainty, and henc e of resolution. Also, how accurate are the methods/instruments used to make such measurements ? None of the work I've read so far addresses these issues, and I would be grateful if someone could post/e-mail potentially useful references. I will post a summary if there is sufficient interest. - - Arun Rao rao@enuxha.asu.edu rao%enuxha.asu.edu@relay.cs.net BITNET: agaxr@asuacvax ------------------------------ Subject: NIPS computer demos From: jbower@bek-mc.caltech.edu (Jim Bower) Date: Tue, 01 Nov 88 13:38:34 -0800 Concerning: Software demonstrations at NIPS Authors presenting papers at NIPS are invited to demo any relevant software either at the meeting itself, or during the post-meeting workshop. The organizers have arranged for several IBMs and SUN workstations to be available. For information on the IBMs contact Scott Kirkpatrick at Kirk@IBM.COM. Two SUN 386i workstations will be available. Each will have a 1/4 cartrage tape drive as well as the standard hard floppies. The machines each have 8 MBytes of memory and color monitors. SUN windows as well as X windows (version 11.3) will be supported. The Caltech neural network simulator GENESIS will be available. For further information on the SUN demos contact: John Uhley (Uhley@Caltech.bitnet) ------------------------------ Subject: Outlets for theoretical work From: INAM000 <INAM%MCGILLB.BITNET@VMA.CC.CMU.EDU> Date: Wed, 19 Oct 88 21:18:00 -0500 Department of Psychology, McGill University, 1205 Avenue Dr. Penfield, Montreal, Quebec, CANADA H3Y 2L2 October 20,1988 Given the recent resurgence of formal analysis of "neural networks" (e.g. White,Gallant,Geman,Hanson,Burr),and the difficulty some people seem to have in finding an appropriate outlet for this work,I would like to remind researchers of the existence of the Journal of Mathematical Psychology.This is an Academic Press Journal that has been in existence for over 20 years,and is quite open to all kinds of mathematical papers in "theoretical" (i.e. mathematical, logical,computational) "psychology" (broadly interpreted). If you want further details regarding the Journal,or feedback about the appropriateness of a particular article,you can contact me by E-mail or telephone (514-398-6128),or contact the Editor directly-J.Townsend,Department of Psychology,Pierce Hall,Rm. 365A,West Lafayette,IND 47907. (E-Mail:KC@BRAZIL.PSYCH.PURDUE.EDU; Telephone: 317-494-6236).The address for manuscript submission is: Journal of Mathematical Psychology, Editorial Office,Seventh Floor, 1250 Sixth Avenue, San Diego, CA 92101. Regards Tony A.A.J.Marley Professor Book Review Editor,Journal of Mathematical Psychology E-MAIL - INAM@MUSICB.MCGILL.CA ------------------------------ Subject: Students for a Better NN Class From: doug@feedme.UUCP (Doug Salot) Organization: Feedme Microsystems, Orange County, CA Date: 28 Oct 88 04:40:53 +0000 I'm currently enrolled in a masters (CS) level neural-net review course, and I offered to solicit the net for suggestions concerning successful approaches and pedagogical tools for such a course. We're currently using the PDP group's epic trilogy and a single ANZA system for experimentation. Both have their problems, but I'd specifically like to hear what has worked for you. Any experience with Grossberg's "Neural Networks and Natural Intelligence"? Other comprehensive books? Good simulation packages? (we've got Suns, Macs, and PCs.) Projects or reading for CS types to understand the properties of non-linear dynamical systems? Good examples for comparing traditional solutions with network solutions? Unfortunately, we're currently studying paradigms that are a couple of years old and whose limitations are immediately apparent (Hopfield, bam, backprop, simple competitive schemes, counterprop). What's showing the most promise in terms of problem solving capability, scalability, and efficiency (here, CS types' only concern with biological feasibility is whether or not their brains can grok the stuff). BTW, if any NN gods or daemons are planning on being in/around Orange County (CA) in the next six weeks or so and feel like charming or dissuading a small group of neural-netaly disenchanted (but latently enthusiastic) students, we'd love to hear from you. Yours for better credit assignment, - -- Doug Salot || doug@feedme.UUCP || ...{zardoz,dhw68k}!feedme!doug ------------------------------ Subject: Suggestions needed... From: spam@sun.soe.UUCP (Crunchy Frog,,,) Date: 24 Oct 88 02:03:16 +0000 I am currently in the "proposal" stage for a undergraduate independant study course. The way I have designed the work so far, I will be doing one credit hour each in Cognitive Psych, Philosophy, and AI. I have had the "traditional" undergrad AI background, and what we did didn't seem to add up very well. In psych I learned about activation networks, and this seemed to me to be the approach for real(tm) A.I. Unfortunately, I don't have much background in this area. My plan is to a) do reading on the physical aspects of brain; what structures exist, what is there "before" anything is learned, and how things are stored. b) I want to read some of the philosophical arguments about intelligence and rational thought, and c) I want to implement a simple network and stick it in an artificial environment. Questions: 1) Any reading list suggestions? 2) Is there any relatively PD software that I can play with relative to this? 3) I'm going to be using Turbo-C with 768K available. Is this insufficient memory? What data structures are used for implementing networks? Should I allow for an arbitrary # of links, or limit the #? 4) Am I reinventing the wheel? Is research so far ahead in this area that there is no way to investigate the "frontiers" of this field w/o billions of years of research, or will I (as I hope) be able to try something novel with the small goal of simulating some tiny aspect of intelligence? 5) What sort of career opportunities are there in this field? Are Universities doing most of this sort of research? Can I do *anything* with my Comp Sci BS degree this May in this area? Am I doomed to writing subroutines in the bowels of IBM mainframes? Thanks for any help - -Roger Gonzalez Please E-mail responses to... - -------- Roger Gonzalez spam@clutx.clarkson.edu Clarkson University spam@clvm.BITNET (315) 268-3748 ------------------------------ Subject: Re: Wanted: info about GENESIS program From: mesard@bbn.com (Wayne Mesard) Date: 23 Oct 88 20:29:09 +0000 >From article <15825@agate.BERKELEY.EDU>, by muffy@violet.berkeley.edu: > I would like some information about the GENESIS neural > simulation program being developed at Cal Tech. It's not a NN simulator. It's a genetic algorithm system for function optimization. Its author is John J. Grefenstette Navy Center for Applied Research in Artificial Intelligence Naval Research Laboratory Washington, D.C. 20375-5000 <gref@AIC.NRL.NAVY.MIL> - -- unsigned *Wayne_Mesard(); I never met a stochastic algorithm I MESARD@BBN.COM didn't like. BBN, Cambridge, MA ------------------------------ End of Neurons Digest *********************