neuron-request@HPLABS.HP.COM ("Neuron-Digest Moderator Peter Marvit") (08/28/89)
Neuron Digest Sunday, 27 Aug 1989 Volume 5 : Issue 35 Today's Topics: Science centre exhibit on neural nets Re: Science centre exhibit on neural nets Schools for AI&Neural-nets Schools for AI&Neural-nets Re: Schools for AI&Neural-nets Schools for Neural Networks and a M.S. Schools for AI/Neural-Nets : Addendum!! Call for papers Tech Report Available: Symbol Grounding Problem ICNC Conference Announcement March 1990 in Germany Fame Problems with the Neural Net as System Model 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: Science centre exhibit on neural nets From: tap@ai.toronto.edu (Tony Plate) Organization: Department of Computer Science, University of Toronto Date: Fri, 21 Jul 89 21:53:45 +0000 [[ Editor's Note: For any readers who have not been to the Toronto Science Center, it is a "must see" if you're in the area. A sort of classy Exploratorium (for those familar with San Francisco). -PM ]] Does anyone have any good ideas for a science center exhibit on neural networks which will give the average visitor some idea of what neural networks are about? I'm helping to design a neural networks exhibit for the Ontario Science Center, a science "museum" in Toronto. The exhibit will probably be interactive - with a few buttons or knobs the visitor can play with. The exhibit should be something which a person with high school education understand. The concepts involved in it should be simple, concrete and interesting - thus things like truth tables are out because they are both abstract and incredibly boring. One problem is that many simple neural networks don't do very surprising or difficult things, e.g. Xor is not really a very difficult problem, it is only interesting to researchers because we know that a perceptron cannot do it. The network has to be simple because the computers available are not very powerful, and we would like it to learn while the visitor is with the exhibit, i.e. in 2 to 5 minutes. They are using Amigas, which seem to be able to do about 1200 link per second for back-propagation training (but maybe that could be improved by doing integer arithmetic and table lookup). We already have a few ideas, but none of them are truly inspirational, so I shan't bore you with them. So if you have or have heard of any good ideas for an interesting and understandable exhibit networks, please mail them to me. If there is enough interest I will post a summary to comp.ai.neural-nets. More context: The exhibit is to be part of a rather large show on Psychology, which will stay on show in Toronto for approximately 8 months, and then move to a museum in the U.S. The organizers want a section on AI, but have decided that this section will consist of one exhibit on neural networks... I'm not getting paid for this, so neither will you if you contribute a great idea. However, I will do my best to make sure you get credited. My connection with the ontario science center is entirely informal, so if you don't want to give away all rights to your ideas, then don't mail them to me. Tony Plate ---------------------- tap@ai.utoronto.ca ----- Department of Computer Science, University of Toronto, 10 Kings College Road, Toronto, Ontario, CANADA M5S 1A4 ------------------------------ Subject: Re: Science centre exhibit on neural nets From: mcvax!ukc!strath-cs!nott-cs!ucl-cs!M.Nigri@uunet.uu.net Date: 27 Jul 89 12:03:26 +0000 From: "Meyer E. Nigri" <M.Nigri@uk.ac.ucl.cs> Tony, One simple demonstration is to show how a nn can recognise patterns. For example, one can present a serie of characters to the nn to be learnt. After the learning phase, one can recover a correct character presenting a corrupted version. With a good graphical interface, I think this example can be interesting. If the graphical interface is a little bit better, a true image could be presented to the nn (like a picture of a car, house, horse, tree etc). 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 | +--------------------------+-----------------------------------------------+ ------------------------------ Subject: Schools for AI&Neural-nets From: ramamurb@turing.cs.rpi.edu (Badrinath Ramamurthy) Organization: RPI CS Dept. Date: Tue, 01 Aug 89 20:19:32 +0000 [[ Editor's Note: Following this message is the set of responses. Remember that "best" is terribly subjective and folks interested in these programs should note the particular biases and strengths of each individual school. For example, while Stanford's CS program is highly rated, participants report that the graduate AI track is very logic oriented -- quite different from teh Cognitiuve Science approach. -PM ]] A friend of mine finished his MS and intends to work for his PhD in Neural Networks and AI. Any suggestions which schools he should be applying to ? You can email me on this address: ramamurb@turing.cs.rpi.edu Thanx in advance. -Badri ------------------------------ Subject: Schools for AI&Neural-nets From: ramamurb@turing.cs.rpi.edu (Badrinath Ramamurthy) Organization: RPI CS Dept. Date: Tue, 08 Aug 89 14:29:31 +0000 Hello friends ! I posted a request for names of schools with good PhD programs for Neural Nets and AI, for a friend of mine. A lot of people were enthusiastic and shared their thoughts. Here is what I received: Many people named the following Schools: Boston U Caltech CMU MIT Rochester U of toronto UCSB Univ of Southern California at LA (USC at LA) UCSD Yale Many people chose to mention names and elaborate, and here I reproduce the text from these responses. Some mention newly formed groups and some give interesting information on literature to look up. ------------------ ------------------------------ ---------- Received: from ucsd.edu by turing.cs.rpi.edu (4.0/1.2-RPI-CS-Dept) id AA16856; Tue, 1 Aug 89 23:06:16 EDT Subject: Re: Schools for AI/Neural nets I am satisfied with UCSD - there are a LOT of people doing work in the area, not only in computer Science but in Cognitive Science, Physics, Linguistics, Economics and Biology (as well as others, no doubt). Course work available include Hecht-Nielsen's full year course, an intro course by Gary Cottrell in Computer Science, seminars on Machine Learning & more. Ongoing seminars include the PDP research group in Cognitive Science, authors & source of the three volume PDP books. Other schools doing interesting things: Boston University & Northeastern have the Center for Adaptive Systems or some such, under Steven Grossberg. The program is new and I don't know much about it, but Grossberg's work (while hard to plow through) is very important. USC has Michael Arbib, Bart Kosko and Christof von der Malsberg. I have heard from a couple of their grad students that contact with them is not easy nor productive. I think very highly of von der Malsberg, however. He is on the right track... Carnegie-Mellon has a major effort. Colorado-Boulder has Mike Mozer & Paul Smolensky. I don't agree with everything they say, but their work is excellent. Mike Jordan is now at M.I.T., he would be excellent to work for in applications of neural networks to robotics problems. Toronto & Rochester have some good people as well. And Stanford now has Rumelhart as well as Mark Gluck, a recent PhD who is extremely prolific. Most of their work is in Psych, or Signal Processing (Bernie Widrow). This should be a good core of schools to look at. I'd recommend getting a hold of recent proceedings from IJCNN, or the last two ICNN conferences, plus the NIPS conference and take a look at papers. If your friend sees something exciting, then he might want to contact that school & see what's up. Lots of research now will pay off in a happy PhD program... Dave DeMers demers@cs.ucsd.edu ----------------------- xx xx xx xx xx xx xx ------------------- >From enorris@gmuvax2.gmu.edu Wed Aug 2 13:20:17 1989 Your friend would do well to consider George Mason University, located near Washington, D.C. There is a novel Ph.D. program in Information Technology which provides a broad base in computer science, software engineering, OR/STAT, etc. The CS Department has a flourishing and well-=funded AI Center with particular interests in Machine Learning. Other CS facultyt are active in neural networks, expert systems, genetic algorithms, natural language processing, etc. Assistantships are available. For firther information write Dr. James Palmer School of Information Technology & Engineering George Mason University Fairfax, VA 22030 (703) 323-2939 Eugene Norris CS Dept, GMU -------------------------------------------------------------- >From honavar@cs.wisc.edu Wed Aug 2 14:05:34 1989 Organization: U of Wisconsin CS Dept CMU, UCSD, Yale, Brown, MIT, UWisc-Madison, Boston U, Stanford, Berkeley, Rochester, UCLA, Maryland, .. He should look at the recent proceedings of IJCAI, NIPS, as well as AI and NNet journals. ---------------------------------------------------------------------- >From heirich@cs.UCSD.EDU Thu Aug 3 00:18:21 1989 Organization: EE/CS Dept. U.C. San Diego You may consider me biased, given my location, but I think the best schools for a graduate program in neural nets are, without a doubt: Univ. Calif. San Diego Cal Tech Univ. Southern Cal. The other significant places would be: U. Toronto; Stanford; U.C. Boulder; Boston U. There are certainly other places, but these should be at the top of any list because of the faculty there. Alan Heirich Comp. Sci. & Eng., Cognitive Science C-014 University of California, San Diego 92093 heirich@cs.ucsd.edu aheirich@ucsd.bitnet -------------------------------------------------------------------- >From ck@rex.cs.tulane.edu Thu Aug 3 02:10:00 1989 At the Computer Science Department of Tulane University we have a PhD program and a group working on Neural Nets. Dr. Koutsougeras ------------------------------------------------------------------------ >From plong@saturn.ucsc.edu Fri Aug 4 14:36:21 1989 Subject: Grad Schools for Neural Nets The University of California at Santa Cruz has an excellent program in Computational Learning Theory, if your friend is interested in studying the theoretical properties of neural nets. David Haussler and Manfred Warmuth, two of the "Four Germans" who wrote a key paper in this area, continue to be central figures, and both are a pleasure to work with. A third faculty member, Dave Helmbold, who has wide ranging interests, has contributed papers recently. Scholars visiting during the summer for joint research projects include Andrzej Ehrenfeucht, Michael Kearns, Nick Littlestone, Rob Schapire and Bob Sloan. The atmosphere here is very relaxed and noncompetative. The campus is truly beautiful, nestled in the redwoods in the hills overlooking Santa Cruz, with much of the campus having a view of the ocean. Computing facilities are excellent, and a vast majority of graduate students receive some sort of assistantship. I strongly recommend UC Santa Cruz for studying the theory of neural nets. Phil Long P.S. Some other schools with a lot of activity in Computational Learning Theory are MIT, Harvard, Technion (Isreal), Penn, Illinois, Illinois-Chicago, and Pittsburgh. ---------------------------------------------------------------------------- Thanks to all the people who gave me this information. In case I receive more responses, I'll post the addenda.(If thats the right word). -Badri ............... Badrinath Ramamurthy ( ramamurb@turing.cs.rpi.edu ) ............... ------------------------------ Subject: Re: Schools for AI&Neural-nets From: Lloyd Lim <iris!lim@UCDAVIS.UCDAVIS.EDU> Organization: U.C. Davis - Department of Electrical Engineering and Computer Science Date: 08 Aug 89 18:22:45 +0000 In previous article ramamurb@turing.cs.rpi.edu (Badrinath Ramamurthy) writes: > > Univ of Southern California at LA (USC at LA) I just thought I'd better clarify so others don't get confused. There is no school that is known as USC at LA. There is the University of Southern California (USC) which is a private school and one of a kind. Then there is the University of California at Los Angeles (UCLA) which is a branch of the University of California system. UCSB, UCSD, and UCD are also UC branches. I suppose that there were responses for both USC and UCLA, thus the confusion. I don't understand how people can get confused. After all, California is THE center of the world. :-) :-D :-) +++ Lloyd Lim Internet: lim@iris.ucdavis.edu (128.120.57.20) Compuserve: 72647,660 US Mail: 146 Lysle Leach Hall, U.C. Davis, Davis, CA 95616 ------------------------------ Subject: Schools for Neural Networks and a M.S. From: mikeoro@hubcap.clemson.edu (Michael K O'Rourke) Organization: Clemson University, Clemson, SC Date: Fri, 11 Aug 89 13:10:21 +0000 I followed with interest the recent discussion on which schools are good for neural nets and AI. I am about to get my B.S. and will enter grad school to get an M.S. in Aug '90. Does everyone pretty much agree that the schools good for the PhD hold true for an M.S. in neural nets? Can anyone suggest other schools that might be better for an M.S. than a PhD? (I am also interesting in Networking and Communications if someone can suggest a school that has these as well as neural nets) Thanks, Michael O'Rourke CLemson University ------------------------------ Subject: Schools for AI/Neural-Nets : Addendum!! From: ramamurb@turing.cs.rpi.edu (Badrinath Ramamurthy) Organization: RPI CS Dept. Date: Tue, 22 Aug 89 13:05:33 +0000 Hi Folks, Here I am, back with a few more responses I received for "Schools for AI/Neural-Nets". Thanks to all people who replied now and then. ======================================================================== >From: patil@a.cs.okstate.edu (Patil Rajendra Bha) Subject: Re: Schools for AI&Neural-nets Date: 11 Aug 89 04:32:47 GMT Organization: Oklahoma State Univ., Stillwater The information about the schools is given in one of the neural network society journal of 1988 some of them are The university of Tennessee , Center for neural engineering Boston University, Center for adaptive systems Brown university, Rhode Island John Hopkins university CALTECH Carnegie Mellon, Psychology Dept, MIT There are many others, last month I posted the same message, I got the same reply what you got, Then I looked into the research center directory and now awating the replies from the universities. I am a graduate student in computer science and willing to go for Ph.D in neural networks. I would appreciate if you could mail the list of universitiesto me. I will let you know about the replies that I will receive. Thank you Patil Rajendra patil@a.cs.okstate.edu ----------------------------------------------------------------------------- Date: Tue, 8 Aug 89 16:54:23 PDT From: ash@cs.UCSD.EDU (Tim Ash) Subject: Re: Schools for AI&Neural-nets The top Neural Net graduate programs are at the following schools: U.C. San Diego U.S.C. C.M.U. Stanford U. Boston U. Some others to consider (not as good as the above): U.C. Boulder C.I.T. University of Toronto Northeastern U. All of the above schools have people who are active and well known in the field. U.C. San Diego is acknowledged to be the strongest school in the field (both in terms of numbers of people, and the variety of academic departments involved in the work). Many of the top researchers at other universities passed through U.C.S.D. at one point or another. Good luck in your friend's search. If you need specific information about U.C.S.D., send me e-mail. Tim Ash (CSE Dept. U.C.S.D.) ash@ucsd.edu ------------------------------------------------------------------------- From: gary@cs.ucsd.edu (Gary Cottrell) How about: University of California San Diego: Researchers in PDP, AI and related fields include: (active research, not department, in parens) Henry Abarbanel (Dynamical systems) Elizabeth Bates (Brain and Language) Rik Belew (Genetic Algs, PDP and AI) Shankar Chatterjee (Vision and simulated annealing) Patricia Churchland (Philosophy of Computational Neuroscience)) Gary Cottrell (PDP) Francis Crick (Neuroscience) Jeff Elman (PDP, NLP and GA) Clark Guest (Optical neurocomputing) Paul Kube (Vision) Marta Kutas (NLP and Neuroscience) David Kirsh (AI) Helen Neville (Brain and Language) Mohan Paturi (Learning theory) Ramachandran (Human Vision) Walt Savitch (NLP) Terry Sejnowski (PDP and Computational Neuroscience) Marty Sereno (PDP and Neuroscience) Hal White (PDP and theory) David Zipser (PDP and Computational Neuroscience) gary cottrell 619-534-6640 Computer Science and Engineering C-014 UCSD, La Jolla, Ca. 92093 gary%cs@ucsd.edu (ARPA) {ucbvax,decvax,akgua,dcdwest}!sdcsvax!gary (USENET) gcottrell@ucsd.edu (BITNET) =================================================================== Hope it helps all souls searching for way into neural-nets/AI ! -Badri ( ramamurb@turing.cs.rpi.edu ) ------------------------------ Subject: Call for papers From: margaux!bouguett@BOULDER.COLORADO.EDU Organization: University of Colorado, Boulder Date: 22 Jul 89 08:15:26 +0000 FIRST MAGHREBIN CONFERENCE ON ARTIFICIAL INTELLIGENCE AND SOFTWARE ENGINEERING Constantine, Algeria, September 24-27, 1989 CALL FOR PAPERS TOPICS The Conference Program will include bith invited and contributed papers. Authors from Maghreb are particulary encouraged to submit. The adressed topics, but not limited to, are : - Algebraic Specification - Program Construction and Proving - Expert Systems - Knowledge and Data Bases - Communication Protocols - Distributed Systems - Object Oriented Programming TERMS OF PRESENTATION OF PAPERS : Papers should be in English, French or Arabic and meet the following requirements : 1- Pages should not number more than 20, including an abstract, tables, figures and references. 2- The papers should be double typed on (A 4) single faced page. 3- The full-name of author (s) and institude and country where the research was conducted should be written on the title page with an abstract of no more than 300 words. 4- Four copies of the papers should be sent to the chaiman of the organizing committee. DEADLINE FOR SUBMISSION OF PAPERS : The closing date for acceptance of papers is 10 August 1989. Those whose papers are accepted will be informed by 4th September 1989. ORGANIZED BY : Laboratory of Knowledge Bases and Distributed Systems Computer Science Institute, Constantine University with the partipation of LRI ORSAY- FRANCE. GUEST SPEAKER : Eric G. Wagner, Research staff member IBM Watson Research Center (USA) CORRESPONDANCE : All correspondance should be adressed to : Dr. BETTAZ Mohamed Institut d'Informatique Universite de Constantine Constantine 25000 ALGERIA Telephone : (213) (4) 69.21.39 Telex : 92436 UNCZL ------------------------------ Subject: Tech Report Available: Symbol Grounding Problem From: harnad@clarity.Princeton.EDU (Stevan Harnad) Date: Sat, 05 Aug 89 01:21:57 -0400 THE SYMBOL GROUNDING PROBLEM Stevan Harnad Department of Psychology Princeton University ABSTRACT: There has been much discussion recently about the scope and limits of purely symbolic models of the mind and about the proper role of connectionism in cognitive modeling. This paper describes the "symbol grounding problem" for a semantically interpretable symbol system: How can its semantic interpretation be made intrinsic to the symbol system, rather than just parasitic on the meanings in our heads? How can the meanings of the meaningless symbol tokens, manipulated solely on the basis of their (arbitrary) shapes, be grounded in anything but other meaningless symbols? The problem is analogous to trying to learn Chinese from a Chinese/Chinese dictionary alone. A candidate solution is sketched: Symbolic representations must be grounded bottom-up in nonsymbolic representations of two kinds: (1) iconic representations, which are analogs of the proximal sensory projections of distal objects and events, and (2) categorical representations, which are learned and innate feature-detectors that pick out the invariant features of object and event categories from their sensory projections. Elementary symbols are the names of these object and event categories, assigned on the basis of their (nonsymbolic) categorical representations. Higher-order (3) symbolic representations, grounded in these elementary symbols, consist of symbol strings describing category membership relations ("An X is a Y that is Z"). Connectionism is one natural candidate for the mechanism that learns the invariant features underlying categorical representations, thereby connecting names to the proximal projections of the distal objects they stand for. In this way connectionism can be seen as a complementary component in a hybrid nonsymbolic/symbolic model of the mind, rather than a rival to purely symbolic modeling. Such a hybrid model would not have an autonomous symbolic "module," however; the symbolic functions would emerge as an intrinsically "dedicated" symbol system as a consequence of the bottom-up grounding of categories' names in their sensory representations. Symbol manipulation would be governed not just by the arbitrary shapes of the symbol tokens, but by the nonarbitrary shapes of the icons and category invariants in which they are grounded. Preprint Available: Stevan Harnad JVNET: harnad@confidence.princeton.edu harnad@princeton.edu srh@flash.bellcore.com harnad@elbereth.rutgers.edu CSNET: harnad%confidence.princeton.edu@relay.cs.net UUCP: harnad@princeton.uucp BITNET: harnad@pucc.bitnet harnad1@umass.bitnet Phone: (609)-921-7771 ------------------------------ Subject: ICNC Conference Announcement March 1990 in Germany From: ECKMILLE@DD0RUD81.BITNET (Rolf Eckmiller, Duesseldorf, FRG) Date: Fri, 18 Aug 89 17:06:00 +0700 CALL FOR PAPERS 8/89 International Conference on: PARALLEL PROCESSING IN NEURAL SYSTEMS AND COMPUTERS (ICNC) - 10th Cybernetics Congress of the DGK - 19. - 21. March, 1990, Heinrich-Heine-Universitaet Duesseldorf (FRG) Organizing Committee: R. Eckmiller (chair), Duesseldorf (FRG) G. Hartmann, Paderborn (FRG) G. Hauske, Muenchen (FRG) C. v.d. Malsburg Los Angeles (USA) W. v. Seelen Bochum (FRG) Conference Language: English Topics: 1) New Concepts in Neuroscience and Computational Neuroscience 2) Massively Parallel Computers (e.g. SUPRENUM, Transputer Systems) 3) Structure and Function of Biological Neural Systems 4) Self Organization versus Programming in Parallel Computers 5) Optical Computers and Molecular Computers 6) Parallel Processing in Artificial Intelligence Activities: *) Invited Lectures by American and European Scientists on the listed Topics *) Oral Presentations (15 + 5 min.) *) Posters Presentations *) Exhibition of Books, Neural Systems, and Computers Invited Speakers include: J.R Barker/Glasgow (UK) G. Carpenter/Boston (USA) J. Feldman/Berkeley (USA) A. Cremers/Dortmund (FRG) H. Haarer/Bayreuth (FRG) K. Fukushima/Osaka (Japan) T. Kohonen/Espoo (Finland) H. Haken/Stuttgart (FRG) D. Psaltis/Pasadena (USA) W. Reichardt/Tuebingen (FRG) U. Trottenberg/Bonn (FRG) Deadline for Submission of Papers:>> 15 November, 1989 << !!!!!!! (4 camera-ready pages per contribution) Publication: Conference Proceedings will be available at the conference as hard cover book (Elsevier Science Publ.) Registration Fees: Before 1 October, 1989 = 150,- DM including Proceedings After 1 October, 1989 = 200,- DM " " Students = 100,- DM " " Students = 50,- DM without Proceedings ICNC-Conference Secretariat: Dr. R. Eckmiller Heinrich-Heine-Universitaet Duesseldorf Div. Biocybernetics Universitaetsstrasse 1 Tel.: (211) 311-5204 D-4000 Duesseldorf (FRG) e-mail: ECKMILLE@DD0RUD81.BITNET Please complete and mail the request of 2nd Announcement to the ICNC-Conference Secretariat. cut here - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - To ICNC-Conference Secretariat Dr. R. Eckmiller Heinrich-Heine-Universitaet Duesseldorf Div. Biocybernetics Universitaetsstrasse 1 D-4000 Duesseldorf (FRG) Request for 2nd Announcement of ICNC 19. - 21. March, 1990, Duesseldorf (FRG) Last Name: ___________________________________________________________________ First Name: __________________________________________________________________ Affiliation: _________________________________________________________________ Street: ______________________________________________________________________ City: ________________________________________________________________________ Country: _____________________________________________________________________ Tel.: e-mail: ______________________________ _________________________________ I intend to participate at the ICNC ( ) as student (yes / no) I intend to submit a contribution on the topic No. ( ) as oral presentation ( ) as poster presentation ( ) as oral paper or poster ( ) Please, send the 2nd Announcement of ICNC to my above address. Date: Sincerely, _________________________ ___________________________________ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ------------------------------ Subject: Fame From: kingsley_morse%01%hp5003@hplabs.hp.com Date: 23 Aug 89 09:20:00 -0800 Want a chance to be famous? Would you like the chance to have your neural net program become a standard of comparison? The major computer companies are cooperating to collect new benchmark programs that will be distributed far and wide. "The prevailing set of synthetic benchmark standards are clearly inadequate to measure today's advanced computer systems," said John Mashey of MIPS Computer Systems. "Small, synthetic benchmarks like Dhrystones and Whetstones can no longer be used to gauge the performance of systems that now take advantage of mainframe and supercomputer design concepts. Today's workstations and servers deliver high performance by utilizing heavy instruction pipelining, multiple execution units working in parallel, large caches, fast memory systems, and optimizing compilers. Some of the existing benchmarks easily fit within the caches of these machines, and the performance results are completely unrealistic. It's like trying to measure the speed of a bullet with a stopwatch." The major computer company cooperative, called SPEC (for Systems Performance Evaluation Cooperative) is collecting programs that are better benchmarks for today's systems. It's members include IBM, DEC, HP, Sun, MIPS, Data General, and Motorola. The programs chosen for this suite of benchmark programs will be an industry standard. If you're interested in sharing the limelight, here's what you should do: 1.) Select a neural net training program and training data whose source code you won't mind putting in the public domain. 2.) Check that your program and training data take 1 to 10 minutes to train on an 8 mips machine. (My understanding is that an HP 835 is about 14 mips, a Sun Sparc station 9 mips, a DEC 11/780 1 mip, and a Decstation 3100 11 mips). 3.) Copy the source code for your neural net program and data to a floppy disc or magnetic tape and mail it by October 15th, 1989 to: Kingsley G. Morse Jr. 1039 Continentals Way, #301 Belmont, CA 94002 I encourage you to submit back-propagation applications written in C or FORTRAN, and to call me at (415) 691-3221 during the day if you have any questions. Sincerely, Kingsley G. Morse Jr. (415) 691-3221 sun.com!hpda!hp5003!01!kingsley_morse ------------------------------ Subject: Problems with the Neural Net as System Model From: fishwick@fish.cis.ufl.edu (Paul Fishwick) Organization: UF CIS Department Date: Mon, 14 Aug 89 14:34:11 +0000 For a pre-print of this article, please send a note with your address to: paulette@bikini.cis.ufl.edu NEURAL NETWORK MODELS IN SIMULATION: A COMPARISON WITH TRADITIONAL MODELING APPROACHES Paul A. Fishwick Department of Computer and Information Science University of Florida Bldg. CSE, Room 301 Gainesville, FL 32611 INTERNET: fishwick@fish.cis.ufl.edu To be presented at: The Winter Simulation Conference, Dec. 1989 ABSTRACT Neural models are enjoying a resurgence in systems research primarily due to a general interest in the connectionist approach to modeling in artificial intelligence and to the availability of faster and cheaper hardware on which neural net simulations can be executed. We have experimented with using a multi-layer neural network model as a simulation model for a basic ballistics model. In an effort to evaluate the efficiency of the neural net implementation for simulation modeling, we have compared its performance with traditional methods for geometric data fitting such as linear regression and surface response methods. Both of the latter approaches are standard features in many statistical software packages. We have found that the neural net model appears to be inadequate in most respects and we hypothesize that accuracy problems arise, primarily, because the neural network model does not capture the system structure characteristic of all physical models. We discuss the experimental procedure, issues and problems, and finally consider possible future research directions. +------------------------------------------------------------------------+ | Prof. Paul A. Fishwick.... INTERNET: fishwick@bikini.cis.ufl.edu | | Dept. of Computer Science. UUCP: gatech!uflorida!fishwick | | Univ. of Florida.......... PHONE: (904)-335-8036 | | Bldg. CSE, Room 301....... FAX is available | | Gainesville, FL 32611..... | +------------------------------------------------------------------------+ ------------------------------ End of Neurons Digest *********************