neuron-request@HPLABS.HPL.HP.COM ("Neuron-Digest Moderator Peter Marvit") (03/04/90)
Neuron Digest Saturday, 3 Mar 1990 Volume 6 : Issue 18 Today's Topics: What is Dr. Stephen Gallant's method Dr. Stephen Gallant's method -- THE ANSWER IS... Chaos in the brain Computational Metabolism on a Connnection Machine and Other Stories... Re: N.N. & CW! Re: Boltzmann v. Cauchy Fujistu "Super Computer" Logic Based Systems Using Neural Nets Two positions Announcement: Workshop on Adapt. Neural Nets & Pattern Recog. NN conference Indiana_Purdue Ft Wayne 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: What is Dr. Stephen Gallant's method From: giant@lindy.Stanford.EDU (Buc Richards) Organization: Stanford University Date: 05 Jan 90 01:25:58 +0000 I've read that Dr. Stephen Gallant has developed a technique for deriving rules from neural networks. Does anyone know what this technique is or have references to papers that describe it? Also is it known if he will be presenting at IJCNN in Washington, DC? Thanks. Buc Richards @ @ Supercomputer Support Staff > Stanford University - ------------------------------ Subject: Dr. Stephen Gallant's method -- THE ANSWER IS... From: giant@lindy.Stanford.EDU (Buc Richards) Organization: Stanford University Date: 25 Jan 90 23:34:00 +0000 The only reference for the method that I received is the following, "Connectionist Expert Systems" in Communications of the ACM February 1988 - Volume 31 - number 2 ------------------------------ Subject: Chaos in the brain From: ssingh@watserv1.waterloo.edu ($anjay "lock-on" $ingh - Indy Studies) Organization: U of Waterloo, Ontario Date: 15 Jan 90 16:19:08 +0000 Here are the some references to chaos-theoretic descriptions of the brain. They are from Neural and Brain Modelling. I also have the Fortran files for the simulation programs contained therein. They are the correct ones. Someone mentioned a few months ago about errors in the book. I hope this proves useful to some of you, given the recent talk about chaos and brains. Chay, T.R. Abnormal discharges and chaos in a neuronal model system. _Biological Cybernetics_. 50, 301-311 Abstract: Using the mathematical model of the pacemaker neuron formulated by Chay, we have investigated the conditions in which a neuron can generate chaotic signals in response to variation in temperature, ionic compositions, chemicals, and the strength of the depolarizing current. Choi, M.Y., and Huberman, B.A. Dynamic Behaviour of nonlinear networks. _Phys. Rev. A_. 28, 1204-1206. Abstract: We study the global dynamics of nonlinear networks made up of synchronous threshold elements. By writing a master equation for the system, we obtain an expression for the time dependence of its activity as a function of parameter values. We show that with both excitatory and inhibatory couplings, a network can display collective behaviour which can be either multiple periodic or deterministic chaotic, a result that appears to be quite general. Grondin, R.O., et. al. Synchronous and Asynchronous Systems of Threshold Elements. _Biological Cybernetics_. 49, 1-7. Abstract: The role of synchronism in systems of threshold elements (such as neural networks) is examined. Some important differences between synchronous and asynchronous systems are outlined. In particular, important restrictions on limit cycles are found in asynchronous systems along with multi-frequency oscillations which do not appear in synchronous systems. The possible role of deterministic chaos in these systems is discussed. Guevara, M.R., Glass, L., Mackey, M.C., Shrier, A. Chaos in Neurobiology. _IEEE Transactions on Systems, Man, and Cybernetics_. 13, 790-798. Abstract: Deterministic mathematical models can give rise to complex aperiodic ("chaotic") dynamics in the abscence of stochastic fluctuations ("noise") in the variables or parameters of the model or in the inputs to the system. We show that chaotic dynamics are expected in nonlinear feedback systems possessing time delays such as are found in recurrent inhibition and from the periodic forcing of neural oscillators. The implications of the possible occurrence of chaotic dynamics for experimental work and mathematical modelling of normal and abnormal function in neurophysiology are mentioned. Holden, A.V., Winlow, W., and Hayden, P.G. The Induction of Periodic and Chaotic Activity in a Molluscan Neurone. _Biological Cybernetics_. 43, 169-173. Abstract: During prolonged exposure to extracellular 4-aminopyridine (4AP) the periodic activity of the somatic membrane of an identified molluscan neurone passes from a repetitive regular discharge of >90 mV amplitude action potentials, through double discharges to <50 mV amplitude oscillations. Return to standard saline causes the growth of parabolic amplitude-modulated oscillations that develop, through chaotic amplitude- modulated oscillations, into regular oscillations. These effects are interpreted in terms of the actions of 4AP on the dynamics of the membrane excitation equations. $anjay "lock-on" $ingh ssingh@watserv1.waterloo.edu "A modern-day warrior, mean mean stride, today's Tom Sawyer, mean mean pride." ------------------------------ Subject: Computational Metabolism on a Connnection Machine and Other Stories... From: marek@iuvax.cs.indiana.edu (Marek Lugowski) Date: 23 Jan 90 16:23:27 +0000 [[ Editor's note: This is a bit afield, and certainly past the presentation date, but I thought readers might be interested in one of the connectionist analogs for a different purpose. -PM ]] Indiana University Computer Science Departamental Colloquium Computational Metabolism on a Connection Machine and Other Stories... --------------------------------------------------------------------- Elisabeth M. Freeman, Eric T. Freeman & Marek W. Lugowski graduate students, Computer Science Department Indiana University Wednesday, 31 January 1990, 7:30 p.m. Ballantine Hall 228 Indiana University campus, Bloomington, Indiana This is work in progress, to be shown at the Artificial Life Workshop II, Santa Fe, February 5-9, 1990. Connection Machine (CM) is a supercomputer for massive parallelism. Computational Metabolism (ComMet) is such computation. ComMet is a tiling where tiles swap places with neighbors or change their state when noticing their neighbors. ComMet is a programmable digital liquid. Reference: Artificial Life, C. Langton, ed., "Computational Metabolism: Towards Biological Geometries for Computing", M. Lugowski, pp. 343-368, Addison-Wesley, Reading, MA: 1989, ISBN 0-201-09356-1/paperbound. Emergent mosaics: - ---------------- This class of ComMet instances arise from generalizing the known ComMet solution of Dijkstra's Dutch Flag problem. This has implications for cryptology and noise-resistant data encodings. We observed deterministic and indeterministic behavior intertwined, apparently a function of geometry. A preliminary computational theory of metaphor: - ---------------------------------------------- We are working on a theory of metaphor as transformations within ComMet. Metaphor is loosely defined as expressing one entity in terms of another, and so it must underlie categorization and perception. We postulate that well-defined elementary events capable of spawning an emergent computation are needed to encode the process of metaphor. We use ComMet to effect this. A generalization of Prisoner's Dilemma (PD) for computational ethics: - --------------------------------------------------------------------- The emergence of cooperation in iterated PD interactions is known. We propose a further generalization of PD into a communication between two potentially complex but not necessarily aware of each other agents. These agents are expressed as initial configurations of ComMet spatially arranged to allow communication through tile propagation and tile state change. Connection Machine (CM) implementation: - -------------------------------------- We will show a video animation of our results, obtained on a 16k-processor CM, including emergent mosaics, thus confirmed after we predicted them theoretically. Our CM program computes in 3 minutes what took 7 days to do on a Lisp Machine. Our output is a 128x128 color pixel map. Our code will run in virtual mode, if need be, with up to 32 ComMet tiles per CM processor, yielding a 2M-tile tiling (over 2 million tiles) on a 64k-processor CM. ------------------------------ Subject: Re: N.N. & CW! From: GMoretti@massey.ac.nz (Giovanni Moretti) Organization: Massey University, Palmerston North, New Zealand Date: 21 Feb 90 02:23:02 +0000 >> Are neural nets suitable for decoding CW (morse code)? Sylvan I won't venture to offer an opinion on the above question (I started reading about NN four days ago) but refer you to a couple of readable articles in BYTE August 1989. The second of these is "Building blocks for Speech" (page 235) which deals with the use of nets to recognise the "B" and "D" consonants, and builds up from there. It's a very readable article about the use of nets on input data that's time varying (there's probably a jargon term for this :-). I've also found a very simple backpropagation program written in Turbo Pascal (about one and a half pages) written by David Parker and included in Dr Dobbs article "Programming Paradigms" by Michael Swaine DDJ October 1989, p112 (listing starts p146). The source is available from SIMTEL-20 in the DDJMAG (I think) directory. These articles aren't exactly high-brow stuff but they're a very good introduction, especially the interview with Dave Parker. I can mail you the Turbo Pascal source if you like. Cheers Giovanni (ZL2BOI) - -- - ------------------------------------------------------------------------------- | GIOVANNI MORETTI, Consultant | EMail: G.Moretti@massey.ac.nz | |Computer Centre, Massey University | Ph 64 63 69099 x8398, FAX 64 63 505607 | | Palmerston North, New Zealand | QUITTERS NEVER WIN, WINNERS NEVER QUIT | - ------------------------------------------------------------------------------- ------------------------------ Subject: Re: Boltzmann v. Cauchy From: munnari!cluster.cs.su.oz.au!ray@uunet.UU.NET Date: Mon, 26 Feb 90 22:32:38 +1100 In Neuron Digest, Volume 6 : Issue 16 (Friday, 23 Feb 1990) andrew@dtg.nsc.com writes ... > Phil Wassermann came to speak at our plant recently, and mentioned that > Cauchy was often superior to Boltzmann in that a large jump out of an > extended local minimum area was more likely, and thus a global minimum > more likely to be successfully found ... A nice property of wandering through an energy landscape by classic simulated annealing is that the relative probabilities of two global states is simply a function of the energies of the two states i.e. it is independent of the path(s) to those states. This is particulary useful for Boltzmann Machines. I don't think this property holds when you use "fast" simulated annealing. Raymond Lister Basser Department of Computer Science University of Sydney NSW 2006 AUSTRALIA Internet: ray@cs.su.oz.AU CSNET: ray%cs.su.oz@RELAY.CS.NET UUCP: {uunet,hplabs,pyramid,mcvax,ukc,nttlab}!munnari!cs.su.oz.AU!ray JANET: munnari!cs.su.oz.AU!ray@ukc (if your lucky) ------------------------------ Subject: Fujistu "Super Computer" From: Mark Dzwonczyk <Mark_Dzwonczyk@qmlink.draper.com> Date: 28 Feb 90 16:19:54 -0800 [[ Editor's note: I took the liberty of attempting to format this article and clean it up a bit. Editorial confusions/comments are marker, as always with [[x]]. -PM ]] Subject: Time: 03:37 PM OFFICE MEMO Fujistu "Super Computer" Date: 2/28/90 Here's an article regarding IJCNN-Wash which mentions the Fujistu machine: There may be some typos; the scanner isn't perfect. Probably needs HNC's OCR system... Mokhoff, N. , "Neural nets making the leap out of lab", Electronic Engineering Times, January 22, 1990, p. 1 & 122. Washington. Neural networks are beginning to make good on their long-promised potential, with individual chips, boards and development tools having poured into the marketplace over the past two years. Last week's International Joint Conference on Neural Networks made it clear that neurocomputers are pleased to take their place as actual partners with digital computers in solving complex problems with unprecedented precision by the year 2000. Among the signposts at IJCNN, aside from the traditional product rollouts, were the debut of a neural ASIC business (see related story, page 122) and a host of presentations on multiprocessor system architectures, most of which are at the prototype stage. Among the developments are: A multiprocessor architecture from Fujitsu Laboratories for connecting up to 1,024 TMS320C30 floating-point digital signal processors to generate a billion connection updates per second. A toroidal lattice architecture implemented with 17 Inmos T800 floating-point Transputers expandable to large-scale networks using digital signal processors from Sharp Corp. A waferscale systolic array comprising more than 900,000 transistors capable of a billion connections per second in a system implementation from Siemens. From SAIC, a concurrent processing architecture that eliminates sequential processing requirementsa throughput limiter [[ sic ]] that can support implementations of both loosely coupled networks and tightly coupled multiple-instruction, multiple-data (MIMD) processors fabricated in VLSI chips. Sandy/8, the TMS320C30 based architecture from Fujitsu Laboratories (Kawasaki, Japan), may be the first such system architecture geared toward use in today's system environment: When full implemented, Sandy/8 will comprise VME boards populated with custom chips that will be scaled replicas of the current prototype board (see photo). Today, one board represents one neuron. The prototype Sandy/8, which will use 256 processing elements when completed in the fall, will eventually be upgraded to the Sandy/10, able to house 1,024 processing elements (PEs). Fujitsu researchers Hideki Kato and Kazuo Asakawa developed the architecture to resolve the thorny problem of how to divide tasks in a multiprocessing system and then assign the divided tasks to PEs without affecting throughput. The basic architecture consists of PEs and "trays" (see diagram on page 1). The trays, which are connected by a ring network that serves as cyclic shift register function as container/routers that allow tasks to be divided and assigned to the PEs. Each tray is implemented in a 20,000-gate Hitachi gate array and is connected to two neighbors. Each PE has a floating-point multiplier, am [[ ??? ]] adder and some local memory (SRAM) to store the weight vectors and program code needed to run the system's backpropagation algorithm. The layers required to implement the various connection patterns are configured in software. "This provides us the flexibility for changing the number of layers to be simulated as required, since the architecture has no physical layer structure," said Asakawa, section manager of the Computer-based System Laboratory. For example, to implement a three-layer perfect connection neural network that has four neurons in the input layer, three in the hidden layer and two in the output layer, four trays and three PEs are needed with no need for any physical layer structure. In addition to the PEs, trays and SRAMs, three other elements implement the architecture: a variable FIFO memory connected to a host Sun-3 computer, short bridges linking the PEs to the ring network of trays and a host common bus through which all of the SRAM associated with the PEs can be accessed through the host. FlFOs and SRAMs are needed to store the vectors, which are used repeatedly in learning cycles to get the best performance and the most accurate results. That's because the bandwidth of the ring network of trays is 67 Mbytes/s much faster than the VME bus connected to the host. Applications other than neuro computing for which the Fujitsu architecture is expected to be useful are 2-D image processing and conventional vector processing. A 2-D version of Sandy for image processing win be ready in two years, using custom VLSI PEs, said senior researcher Kato. Among the ready-for-market products shown was the most commercial, turnkey, end-user neural network product. Hecht-Nielsen Neurocomputer (San Diego) introduced IDEPT, the Image Document Entry Processing Terminal, for recognizing a wide variety of images. Armed with Oscar, HNC's proprietary neural network recognition software, IDEPT is claimed to "achieve on a digit-to-digit basis a greater accuracy m handwritten numbers and alphabetic characters than any other image character recognition system," said Robert Hecht-Nielsen, HNC's chairman. For $39,500, the IDEPT workstation includes an AT-compatible personal computer, am HNC Anza neurocomputing coprocessor, an optical scanner with document feeder, a color VGA monitor and Oscar. IDEPT can operate as a standalone unit or m a networked configuration, where it can recognize characters on forms that have been stored on a remote image data system or scanned at another location. HNC also incorporated its General Applications Architecture (GAA) into NetsetII, a software package that allows users to define and create applications without programming. GAA elements are represented by icons on the screen and include the following: file processor, data processor, pipe processor, display processor and a neural network processor that allows access to all 19 neural network paradigms supplied by HNC. "GAA is the first modular CASE environment built by a neural network company," said Hecht-Nielsen. Meanwhile, Lucid Inc. (Menlo Park, Calif.) said it will supply Plexi, a software-based neural net system for Unix platforms. Offered for the Sun-3, Sun-4 and Sparcstation in the summer, Plexi is a general development tool for neural network simulations that was developed by Plexi Software Inc. (San Francisco). Lucid, a leading supplier of Common Lisp software products, will distribute Plexi development and delivery products and plans to port Plexi to all Lucid Common Lisp platforms. The product is menu driven, and networks can be con structured and modified via mouse actions using the Network Editor and Pattern Editor. Several conventional neural network paradigms Hopfield memory, back propagation and competitionare provided with Plexi. [[ sic ]] ------------------------------ Subject: Logic Based Systems Using Neural Nets From: arun1@mars.njit.edu (arun maskara spec lec cis) Date: Thu, 01 Mar 90 14:12:46 -0500 "Logic Based Systems Using Neural Nets" Since last 9 months, I have been thinking about the idea of implementing an inference engine using Neural Network. I have written simulation program which implements Inference engine, and works on simple set of rules. I want to get feedback from other people who have thought of this idea or are working on something related. Thank You all. Arun Maskara arun1@mars.njit.edu ------------------------------ Subject: Two positions From: marwan@extro.ucc.su.oz.au (Marwan Jabri) Organization: University Computing Service, Uni. of Sydney, Australia. Date: 24 Feb 90 01:52:59 +0000 [[ Editor's note: I assume the salaries listed are in Australian dollars; convert accordingly. -PM ]] Sydney University Electrical Engineering Systems Engineering and Design Automation Laboratory (Two Positions) - --------------------------------------------------------------- Research Fellow Reference No. 08/01 Microelectronic Implementation of Neural Networks based Devices for the Analysis and Classification of Medical Signals Applications are invited from enthusiastic persons to work on advanced neural network application project in the medical area. The project is being funded jointly by the Australian Goverment and a high-technology manufacturer of medical products. The project is the research and development of different architectures of networks to be implemented on ASIC's. The chips are to be used for the analysis and classification of medical signals. The successful applicant is expected to play a leading role in system specification, design and implementation, functional and circuit level simulation. Applicants should have an electrical engineering degree or equivalent, and either a PhD degree or a substantial experience in a related field. The appointees may apply for enrollment towards a postgraduate degree (part-time). Preference will be given to applicants who have experience in artificial neural networks, MOS analog or digital integrated circuit design. The appointment is originally for one year with possibility of renewal. Salary range according to qualifications. - --------------------------------------------------------------- Professional Assistant Grade I/II Reference No. 08/04 Microelectronic Implementation of Neural Networks based Devices for the Analysis and Classification of Medical Signals Applications are invited from enthusiastic persons to work on advanced neural network application project in the medical area. The project is being funded jointly by the Australian Goverment and a high-technology manufacturer of medical products. The project is the research and development of different architectures of networks to be implemented on ASIC's. The chips are to be used for the analysis and classification of medical signals. The successful applicant is expected to contribute to architecture design and implementation, and to carry out MOS circuit design, modelling, and simulation. Applicants should have an electrical engineering degree or equivalent. Preference will be given to applicants who have experience in artificial neural networks, MOS analog or digital integrated circuit design. The appointees may apply for enrollment towards a postgraduate degree (part-time). The appointment is originally for one year with possibility of renewal. Salary range according to qualifications. - --------------------------------------------------------------- Salaries: Research Fellow $32,197-$41,841 p.a. Professional Assistant Grade II $33,842-$35,928 Professional Assistant Grade I $23,878-$32,053 Method of application: Applications, quoting ref. no. and including curriculum vitea, list of publications and the names, addresses and Fax nos, of three referees, to the Registrar, Staff Office, University of Sydney, NSW 2006, from whom general information is available. The University reserves the right not to proceed with any appointment for financial or other reasons. No smoking in the workplace is University policy --------------------------------------------------------------- For further information please contact: Dr M.A. Jabri Sydney University Electrical Engineering NSW 2006 Australia Tel: +61-2-692 2240 Fax: +61-2-692 3847 Email: marwan@ee.su.oz.au Marwan Jabri E-mail: marwan@ee.su.oz Systems Engineering and Design Automation Laboratory Fax: (+61-2) 692 3847 Sydney University Electrical Engineering NSW 2006 Australia ------------------------------ Subject: Announcement: Workshop on Adapt. Neural Nets & Pattern Recog. From: flynn@pixel.cps.msu.edu (Patrick J. Flynn) Date: Tue, 27 Feb 90 08:24:06 -0500 Workshop on Artificial Neural Networks & Pattern Recognition Sponsored by The International Association for Pattern Recognition (IAPR) Sands Hotel Atlantic City, New Jersey June 17, 1990 Recent developments in artificial neural networks (ANN's) have caused a great deal of excitement in the academic, industrial, and defense communities. Current ANN research owes much to several decades of work in statistical pattern recognition (SPR); indeed, many fundamental concepts from SPR have recently found new life as research topics when placed into the framework of an ANN model. The aim of this one-day workshop is to provide a forum for itneraction between the leading researchers from the SPR and ANN fields. As pattern recognition practioners, we seek to address the following issues: **In what ways do artificial neural networks differ from the well-known paradigms of statistical pattern recognition? Are there concepts in ANN for which no counterpart in SPR exists (and vice versa?) **What benefits can come out of interaction between ANN and SPR researchers? **What advantages, if any, does ANN techniques have over SPR methods in dealing with real world problems such as object recognition, pattern classification, and visual environment learning? Tentative Program 8:00 Registration 8:30 Issues in ANN and SPR, Laveen Kanal, University of Maryland 9:15 Links Between ANN's & SPR, Paul Werbos, National Science Foundation 10:00 Coffee Break 10:30 Generalization & Discovery in Adaptive Pattern Recognition, Y. Pao, Case Western Reserve University 11:15 Character Recognition, Henry Baird, AT&T Bell Labs 12:00 LUNCH 1:30 Target Recognition, Steven Rogers, U.S. Air Force 2:15 Connectionist Models for Speech Recognition, Renato DeMori, McGill University 3:00 Coffee Break 3:30 Panel Discussion, Moderators: Anil Jain, Michigan State University & Ishwar Sethi, Wayne State University Registration Information: Advance Registration (by 5/15/90): $100 Late Registration: $120 Contact: Ms. Cathy Davison (Workshop on ANN and PR) Department of Computer Science, A-714 Wells Hall Michigan State University, East Lansing, MI 48824 Tel. (517)355-5218, email: davison@cps.msu.edu, FAX: (517)336-1061 ------------------------------ Subject: NN conference Indiana_Purdue Ft Wayne From: Samir Sayegh <sayegh@ed.ecn.purdue.edu> Date: Thu, 01 Mar 90 18:47:15 -0500 Third Conference on Neural Networks and Parallel Distributed Processing Indiana-Purdue University A conference on NN and PDP will be held April 12, 13 and 14, 1990 on th common campus of Indiana and Purdue University at Ft Wayne. The emphasis of this conference will be Vision and Robotics although all contributions are w are welcome. People from the Midwest are particularly encouraged to attend and contribute especially since the "major" NN conferences seem to oscillate between the East and West Coast! Send abstracts and inquiries to: Dr. Samir Sayegh Physics Department Indiana Purdue University Ft Wayne, IN 46805 email: sayegh@ed.ecn.purdue.edu sayegh@ipfwcvax.bitnet FAX : (219) 481-6800 Voice: (219) 481-6157 ------------------------------ End of Neuron Digest [Volume 6 Issue 18] ****************************************