NEURON-Request@ti-csl.csc.ti.COM (NEURON-Digest moderator Michael Gately) (02/12/88)
NEURON Digest Thu Feb 11 17:26:23 CST 1988 Volume 3 / Issue 5 Today's Topics: neural-net Learning analogies in neural nets Adaline learning rule Who's doing What in neural net research? A short definition of Genetic Algorithms Re: Cognitive System using Genetic Algo Stanford Adaptive Networks Colloquium Preprint available on adaptive network models of human learning ---------------------------------------------------------------------- Date: Wed, 03 Feb 88 12:46:07 EST From: wall <WALL%SBCCVM.BITNET@mitvma.mit.edu> Subject: Re: NEURON Digest - V3 #3 does anyone know of any work being done in the use of more biologically motivated models of neural nets? lets face it, this PDP stuff is getting pretty far afield of real neurons, and i just wondered wether recent research in neurology is being put to use in AI systems. any help on this would be greatly appreciated. 10Q in advance wallace marshall WALL@SBCCVM.BITNET ------------------------------ Date: 5 Feb 88 04:39:58 GMT From: Francis Kam <nuchat!uhnix1!cosc2mi@uunet.uu.net> Subject: neural-net I am working on the learning aspects of the neural net model in computing and would like to know what's happening in the rest of the neural net community in the following areas: 1) neural net models 2) neural net learning rules 3) experimental (analog, digital, optical) results of any kind with figures; 4) neural net machines (commercial, experimental, any kind); 5) any technical reports in these areas; For information exchange and discussion purpose, please send mail to mkkam@houston.edu. Thank you. ------------------------------ Date: 9 Feb 88 04:49:12 GMT From: Francis Kam <nuchat!uhnix1!cosc2mi@uunet.uu.net> Subject: Who's doing What in neural net research? I posted a request for neural net research information which was considered too general. My objective is to gather information of technical reports, papers, on-going research projects, experimental results, etc. I think information like this will facilitate further research in neural net or related computation for the research community as a whole. If there're already sources of information like these, please let me know. Information that I collected from responses of this article will be posted onto the bulletin board or sent by email. To be specific, interested(to me) areas are: (not meant to be comprehensive) 1) neural net model (or PDP model) as a general model of parallel computation -- validity, applicable problem domains, performance analysis, levels and structures, etc.; 2) neural machines -- experimental and commercial; I'm aware of Mark III Mark IV by TRW electronics (details not known); any work on hypercubes and Connection Machine, etc.; in communication intensive model like neural net, is there any successful simulation done on an Ethernet of Sun's, for example; 3) neural net programming environment -- language of neuronal computation description, network layout description, monitoring, software development environment etc.; 4) learning in neural net -- learning rules, memory capacity, behavior such as forgetting and re-learn, etc.; 5) applications -- optimization problems, ai applications, any others to show the usefulness of this model. Please send your responses to: CSNET : mkkam@houston.edu USmail: Francis Kam PGH 550, Department of Computer Science University of Houston 4800 Calhoun, Houston, TX 77004. ------------------------------ Date: 8 Feb 88 15:17:00 GMT From: merrill@iuvax.cs.indiana.edu Subject: Learning analogies in neural nets I would be interested in hearing from anyone who has been working on analogy in neural networks, and, more specifically, learning analogies in that context. Please E-mail any responses, and I'll post a summary to the network. --- John Merrill ARPA: merrill@iuvax.cs.indiana.edu UUCP: {pyramid, ihnp4, pur-ee, rutgers}!iuvax!merrill ------------------------------ Date: 9 Feb 88 18:47:03 GMT From: Thomas E Burns <macintos@ee.ecn.purdue.edu> Subject: Adaline learning rule Does anyone know of any studies on the learning/recal capacity of the Adaline learning Rule? Also does anyone know of the learning rule with the highest learn/recal capacity? I would appreciate any information. Will at Purdue ------------------------------ Date: Fri, 5 Feb 88 10:08:02 PST From: Rik Belew <rik@sdcsvax.ucsd.edu> Subject: A short definition of Genetic Algorithms Mark Goldfain asks: Would someone do me a favor and post or email a short definition of the term "Genetic Learning Algorithm" or "Genetic Algorithm" ? I feel like Genetic Algorithms has two, not quite distinct meanings these days. First, there is a particular (class of) algorithms developed by John Holland and his students. This GA(1) has at its most distinctive feature the "cross-over" operator, which Holland has gone to some effort to characterize analytically. Then there is a broader class GA(2) of genetic algorithms (sometimes also called "simulated evolution") that bear some loose resemblence to population genetics. These date back to at least Fogel, Owen and Walsh (1966). Generally, these algorithms make use of only a "mutation" operator. The complication comes with work like Ackley's thesis (CMU, 1987) which refers to Holland's GA(1), but which is most accurately described as a GA(2). Richard K. Belew rik@cs.ucsd.edu Computer Science & Engr. Dept. (C-014) Univ. Calif - San Diego San Diego, CA 92093 ------------------------------ Date: 5 Feb 88 18:22:21 GMT From: g451252772ea@deneb.ucdavis.edu Subject: Re: Cognitive System using Genetic Algo I offer definitions by (1) aspersion (2) my broad characterization (3) one of J Holland's shortest canonical characterizations and (4) application. (1) GA are anything J Holland and/or his students say they are. (But this _is_ an aspersion on a rich, subtle and creative synthesis of formal systems and evolutionary dynamics.) (2) Broadly, GA are an optimization method for complex (multi-peaked, multi- dimensional, ill-defined) fitness functions. They reliably avoid local max/min, and the search time is much less than random search would require. Production rules are employed, but only as mappings from bit-strings (with wild-cards) to other bit strings, or to system outputs. System inputs are represented as bitstrings. The rules are used stochastically, and in parallel (at least conceptually; I understand several folk are doing implementations, too). A pretty good context paper for perspective (tho weak on the definition of GA!) is the Nature review 'New optimization methods from physics and biology' (9/17/87, pp.215-19). The author discusses neural nets, simulated annealing, and one example of GA, all applied to the TSP, but comments that "... a thorough comparason ... _would be_ very interesting" (my emphasis). (3) J. Holland, "Genetic algorithms and adaptation", pp. 317-33 in ADAPTIVE CONTROL OF ILL-DEFINED SYSTEMS, 1984, Ed. O. Selfridge, E. Rissland, M. A. Arbib. Page 319 has: "In brief, and very roughly, a genetic algorithm can be looked upon as a sampling procedure that draws samples from the set C; each sample drawn has a value, the fitness of the corresponding genotype. >From this point of view the population of individuals at any time t, call it B(t), is a _set_ of samples drawn from C. The genetic algo- rithm observes the fitnesses of the individuals in B(t) and uses this information to generate and test a new set of individuals, B(t+1). As we will soon see in detail, the genetic algorithm uses the familiar "reproduction according to fitness" in combination with crossing over (and other genetic operators) to generate the new individuals. This process progressively biases the sampling pro- cedure toward the use of _combinations_ of alleles associated with above-average fitness. Surprisingly, in a population of size M, the algorithm effectively exploits some multiple of M^3 combinations in exploring C. (We shall soon see how this happens.) For populations of more than a few individuals this number, M^3, is vastly greater than the total number of alleles in the population. The correspond- ing speedup in the rate of searching C, a property called _implicit parallelism_, makes possible very high rates of adaptation. Moreover, because a genetic algorithm uses a distributed database (the popu- lation) to generate new samples, it is all but immune to some of the difficulties -- false peaks, discontinuities, high-dimensionality, etc. -- that commonly attend complex problems." Well, _I_ shall soon close here, but first the few examples of applications that I know of (the situation reminds me of the joke about the two rubes visiting New York for the first time, getting off the bus with all of $2.50. What to do? One takes the money, disappears into a drugstore and reappears having bought a box of Tampax. Quoth he, "With tampax, you can do _anything_!) Anyway: o As noted, the TSP is a canonical candidate. o A student of Holland has implemented a control algorithm for a gas pipe-line center, which monitors and adaptively controls flow rates based on cyclic usages and arbitrary, even ephemeral, constraints. o Of course, some students have done some real (biological) population genetics studies, which I note are a tad more plausible than the usual haploid, deterministic equations. o Byte mag. has run a few articles, e.g. 'Predicting International Events' and 'A bit-mapped Classifier' (both 10/86). o Artificial animals are being modelled in artificial worlds. (When will the Vivarium let some their animated blimps ("fish") be so programmed?) Finally, I noted above that the production rules take system inputs as bit-strings. This representation allows for induction, and opens up a large realm of cognitive science issues, addressed by Holland et al in their newish book, INDUCTION. Hope this helps. I really would like to hear about other application areas; pragmatic issues are still unclear in my mind also, but as apparent, the GA model has intrinsic appeal. Ron Goldthwaite / UC Davis, Psychology and Animal Behavior 'Economics is a branch of ethics, pretending to be a science; ethology is a science, pretending relevance to ethics.' ------------------------------ ------------------------------ Date: Mon, 1 Feb 88 14:13:38 EST From: "Stephen J. Hanson" <jose@tractatus.bellcore.com> Subject: please post Connectionist Modeling and Brain Function: The Developing Interface February 25-26, 1988 Princeton University Lewis Thomas Auditorium This symposium explores the interface between connectionist modeling and neuroscience by bringing together pairs of collaborating speakers or researchers working on related problems. The speakers will consider the current state and future prospects of four fields in which convergence between experimental and computational approaches is developing rapidly. Thursday Friday Associative Memory and Learning Sensory Development and Plasticity 9:00 am 9:00 am Introductory Remarks Preliminaries Professor G. A. Miller Announcements 9:15 am 9:15 am Olfactory Process and Associative Role of Neural Activity in the Memory: Cellular and Modeling Development of the Central Visual Studies System: Phenomena, Possible Mechanism and a Model Professor A. Gelperin Professor Michael P. Stryker AT&T Bell Laboratories University of California, San Francisco Princeton University 10:30 am 10:30 am Simple Neural Models of Towards an Organizing Principle for a Classical Conditioning Perceptual Network Dr. G. Tesauro Dr. R. Linsker, M.D., Ph.D. Center for Complex Systems Research IBM Watson Research Lab Noon-Lunch Noon-Lunch 1:30 pm 1:30 pm Brain Rhythms and Network Memories: Biological Constraints on a Dynamic I. Rhythms Drive Synaptic Change Network: Somatosensory Nervous System Professor G. Lynch Dr. T. Allard University of California, Irvine University of California, San Francisco 3:00 pm 3:00 pm Brain Rhythms and Network Memories: Computer Simulation of Representational II. Rhythms Encode Memory Plasticity in Somatosensory Cortical Hierarchies Maps Professor R. Granger Professor Leif H. Finkel University of California, Irvine Rockefeller University The Neuroscience Institute 4:30 pm General Discussion 4:30 pm General Discussion 5:30 pm Reception 5:30 pm Reception Green Hall, Langfeld Lounge Green Hall, Langfeld Lounge Organizers Sponsored by Stephen J. Hanson Bellcore & Department of Psychology Princeton U. Cognitive Science Laboratory Carl R. Olson Princeton U. Human Information Processing Group George A. Miller, Princeton U. (new page) Connectionist Modeling and Brain Function: The Developing Interface February 25-26, 1988 Princeton University Lewis Thomas Auditorium Travel Information Princeton is located in central New Jersey, approximately 50 miles southwest of New York City and 45 miles northest of Philadelphia. To reach Princeton by public transportation, one usually travels through one of these cities. We recommend the following routes: By Car >From NEW YORK - - New Jersey Turnpike to Exit #9, New Brunswick; Route 18 West (approximately 1 mile) to U.S. Route #1 South, Trenton. From PHILADELPHIA - - Interstate 95 to U.S. Route #1 North. From Washington - - New Jersey Turnpike to Exit #8, Hightstown; Route 571. Princeton University is located one mile west of U.S. Route #1. It can be reached via Washington Road, which crosses U.S. Route #1 at the Penns Neck Intersection. By Train Take Amtrak or New Jersey Transit train to Princeton Junction, from which you can ride the shuttle train (known locally as the "Dinky") into Princeton. Please consult the Campus Map below for directions on walking to Lewis Thomas Hall from the Dinky Station. For any further information concerning the conference please contact our conference planner: Ms. Shari Landes Psychology Department Princeton University, 08544 Phone: 609-452-4663 Elec. Mail: shari@mind.princeton.edu ------------------------------ Date: Thu, 4 Feb 88 09:30:44 PST From: Mark Gluck <netlist@psych.stanford.edu> Subject: Stanford Adaptive Networks Colloquium Stanford University Interdisciplinary Colloquium Series: ADAPTIVE NETWORKS AND THEIR APPLICATIONS ************************************************************************** Feb. 9th (Tuesday, 3:15pm) Tom Landauer "Trying to Teach a Backpropagation Network Bellcore to Recognize Elements of Continuous Speech." ************************************************************************** Abstract We have been trying to get a backpropagation network to learn to spot elementary speech sounds (usually "demisyllables") in continuous speech. The input is typically a 150 msec moving window of sound preprocessed to a spectrotemporal representation resembling the transform imposed by the ear. Output nodes represent speech sounds; their goal states are determined by what speech sounds a human recognizes in the same sound segment. I will describe a variety of small experiments on training regimens and parameters, feedback variations and tests of generalization. . . . . Format: Tea will be served 15 minutes prior to the talk, outside the lecture hall. The talks (including discussion) last about one hour. Following the talk, there will be a reception in the fourth floor lounge of the Psychology Dept. Location: Room 380-380W, which can be reached through the lower level between the Psychology and Mathematical Sciences buildings. Technical Level: These talks will be technically oriented and are intended for persons actively working in related areas. They are not intended for the newcomer seeking general introductory material. Comming up next: Yann Le Cun (Univ. of Toronto) on Friday, Feb. 12th at 1:15pm in room 420-050. Co-sponsored by the Depts. of Psychology and Electrical Engineering ------------------------------ Date: Tue, 9 Feb 88 09:37:33 PST From: Mark Gluck <netlist@psych.stanford.edu> Subject: Stanford Adaptive Networks Colloquium Stanford University Interdisciplinary Colloquium Series: Adaptive Networks and their Applications NOTE: TWO TALKS THIS WEEK Tom Landauer & Yann Le Cun TODAY: Feb. 9th (Tuesday, 3:15pm) Tom Landauer "Trying to Teach a Backpropagation Network Bellcore to Recognize Elements of Continuous Speech." ************************************************************************** FRIDAY: Feb. 12th (Friday, 1:15pm) -- note time Yann Le Cun "Pseudo-Newton and Other Variations of Dept. of Computer Science, Backpropagation" University of Toronto Candada M5S 1A4 Abstract Among all the learning procedures for connectionist networks, the back-propagation algorithm (BP) is probably the most widely used. However, little is known about its convergence properties. We propose a new theoretical framework for deriving the BP based on the Langrangian formalism. This method is similar to some of the methods used in optimal control theory. We derive some variations of the basic procedure, including a pseudo-Newton method that uses the second derivative of the cost function. We also present some results involving networks with constrained weights. It is shown that this technique can be used for putting some a priori knowledge into the network in order to improve the generalization. . . . . Format: Tea will be served 15 minutes prior to the talk, outside the lecture hall. The talks (including discussion) last about one hour. Following the talk, there will be a reception in the fourth floor lounge of the Psychology Dept. Location: Room 380-380W, which can be reached through the lower level between the Psychology and Mathematical Sciences buildings. The Friday talk will be next door in 050. Technical Level: These talks will be technically oriented and are intended for persons actively working in related areas. They are not intended for the newcomer seeking general introductory material. Information: For additional information, contact Mark Gluck (gluck@psych.stanford.edu) 415-725-2434. Comming up next: Mar. 9th (Wednesday, 3:45pm) Jeffrey Elman "Processing Language Without Symbols? Dept. of Linguistics, A Connectionist Approach" U.C., San Diego * * * Co-Sponsored by: Departments of Electrical Engineering (B. Widrow) and Psychology (D. Rumelhart, M. Gluck), Stanford Univ. ------------------------------ Date: Fri, 5 Feb 88 06:25:22 PST From: Mark Gluck <gluck@psych.stanford.edu> Subject: Preprint available on adaptive network models of human learning Copies of the following preprint are available to those who might be interested: Gluck, M. A., & Bower, G. H. (in press). Evaluating an adaptive network model of human learning, Journal of Memory and Language. ------------------------------ Abstract -------- This paper explores the promise of simple adaptive networks as models of human learning. The least-mean-squares (LMS) learning rule of networks corresponds to the Rescorla-Wagner model of Pavlovian conditioning, suggesting interesting parallels in human and animal learning. We review three experiments in which subjects learned to classify patients according to symptoms which had differing correlations with two diseases. The LMS network model predicted the results of these experiments, comparing somewhat favorably with several competing learning models. We then extended the network model to deal with some attentional effects in human discrimination learning, wherein cue weight reflects attention to a cue. We further extended the model to include conjunctive features, enabling it to approximate classic results of the difficulty ordering oflearning differing types of classifications. Despite the well-known limitations of one-layer network models, we nevertheless promote their use as benchmark models because of their explanatory power, simplicity, aesthetic grace, and approximation, in many circumstances, to multilayer network models. The successes of a simple model suggest greater accuracy of the LMS algorithm against other learning rules, while its failures inform and constrain the class of more complex models needed to explain complex results. . . . . For copies, netmail to gluck@psych.stanford.edu or USmail to: Mark Gluck Dept. of Psychology Bldg. 420; Jordan Hall Stanford Univ. Stanford, CA 94305 (415) 725-2434 ------------------------------ End of NEURON-Digest ********************