neuron-request@HPLABS.HP.COM (Neuron-Digest Moderator Peter Marvit) (12/12/88)
Neuron Digest Sunday, 11 Dec 1988 Volume 4 : Issue 32 Today's Topics: Tech report abstracts NN training program at UCSD No hidden neurons, BP vs Perceptrons. Report available. Stefan Shrier on Abduction Machines for Grammars Stanford Adaptive Networks Colloquium TR from ICSI on "Knowledge-Intensive Recruitment Learning" INTERFACE Call for Commentators and/or Original Contributions. [[ Editor's Note: As keeping with reader requests, this issue is strictly tech reports, and announcements. "Discussions" next issue. -PM ]] Send submissions, questions, address maintenance and requests for old issues to "neuron-request@hplabs.hp.com" or "{any backbone,uunet}!hplabs!neuron-request" ------------------------------------------------------------ Subject: Tech report abstracts From: honavar@cs.wisc.edu (A Buggy AI Program) Date: Wed, 30 Nov 88 17:23:01 -0600 The following technical reports are now available. Requests for copies may be sent to: Linda McConnell Technical reports librarian Computer Sciences Department University of Wisconsin-Madison 1210 W. Dayton St. Madison, WI 53706. USA. or by e-mail, to: linda@shorty.cs.wisc.edu PLEASE DO NOT REPLY TO THIS MESSAGE, BUT WRITE TO THE TECH REPORTS LIBRARIAN FOR COPIES. -- Vasant Computer Sciences TR 793 (also in the proceedings of the 1988 connectionist models summer school, (ed) Sejnowski, Hinton, and Touretzky, Morgan Kauffmann, San Mateo, CA) A NETWORK OF NEURON-LIKE UNITS THAT LEARNS TO PERCEIVE BY GENERATION AS WELL AS REWEIGHTING OF ITS LINKS Vasant Honavar and Leonard Uhr Computer Sciences Department University of Wisconsin-Madison Madison, WI 53706. U.S.A. Abstract Learning in connectionist models typically involves the modif- ication of weights associated with the links between neuron-like units; but the topology of the network does not change. This paper describes a new connectionist learning mechanism for generation in a network of neuron-like elements that enables the network to modify its own topology by growing links and recruiting units as needed (possibly from a pool of available units). A combination of generation and reweighting of links, and appropriate brain-like constraints on network topology, together with regulatory mechan- isms and neuronal structures that monitor the network's performance that enable the network to decide when to generate, is shown capa- ble of discovering, through feedback-aided learning, substantially more powerful, and potentially more practical, networks for percep- tual recognition than those obtained through reweighting alone. The recognition cones model of perception (Uhr1972, Hona- var1987, Uhr1987) is used to demonstrate the feasibility of the approach. Results of simulations of carefully pre-designed recog- nition cones illustrate the usefulness of brain-like topological constraints such as near-neighbor connectivity and converging- diverging heterarchies for the perception of complex objects (such as houses) from digitized TV images. In addition, preliminary results indicate that brain-structured recognition cone networks can successfully learn to recognize simple patterns (such as letters of the alphabet, drawings of objects like cups and apples), using generation-discovery as well as reweighting, whereas systems that attempt to learn using reweighting alone fail to learn. ----------------------------------------------------- Computer Sciences TR 805 Experimental Results Indicate that Generation, Local Receptive Fields and Global Convergence Improve Perceptual Learning in Connectionist Networks Vasant Honavar and Leonard Uhr Computer Sciences Department University of Wisconsin-Madison Abstract This paper presents and compares results for three types of connectionist networks: [A] Multi-layered converging networks of neuron-like units, with each unit connected to a small randomly chosen subset of units in the adjacent layers, that learn by re-weighting of their links; [B] Networks of neuron-like units structured into successively larger modules under brain-like topological constraints (such as layered, converging-diverging heterarchies and local recep- tive fields) that learn by re-weighting of their links; [C] Networks with brain-like structures that learn by generation- discovery, which involves the growth of links and recruiting of units in addition to re-weighting of links. Preliminary empirical results from simulation of these net- works for perceptual recognition tasks show large improvements in learning from using brain-like structures (e.g., local receptive fields, global convergence) over networks that lack such structure; further substantial improvements in learning result from the use of generation in addition to reweighting of links. We examine some of the implications of these results for perceptual learning in con- nectionist networks. ------------------------------ Subject: NN training program at UCSD From: elman@amos.ling.ucsd.edu (Jeff Elman) Date: Wed, 30 Nov 88 19:47:48 -0800 RESEARCH AND TRAINING PROGRAM IN NEURAL MODELLING FOR DEVELOPMENTAL PSYCHOLOGISTS University of California, San Diego The Center for Research in Language at UCSD has just obtained a pilot grant from the John D. and Catherine T. MacArthur Foundation, to provide 5 - 10 developmental psychologists at any level (dissertation students through senior investigators) with short-term training in neural computation. The program has two goals: (1) To encourage developmental psychologists in target interest areas (speech, language, early visual-motor and cognitive development, future oriented processes) to begin making use of connectionist modelling as a tool for evaluating theories of learning and change; (2) To encourage greater use of realistic developmental data in the connectionist enterprise. Our experience at UCSD suggests that a well-prepared and computer literate developmental psychologist can learn to make productive use of neural modelling techniques in a relatively short period of time, i.e. 2 weeks to 3 months, depending on level of interest and prior experience. Appli- cants may request training periods in this range at any point from 9/89 through 8/90. Depending on the trainee's needs and resources, we will provide (1) lodging at UCSD, (2) travel (in some cases), (3) access to SUN and VAX works- tations with all necessary software, and (4) hourly services of an individual programmer/tutor who will supervise the trainee's progress through self-paced learning materials while assisting in the implementation of the trainee's pro- posed developmental project. Trainees are also welcome to attend seminars and workshops, and to consult with the rela- tively large number of faculty involved in connectionist modelling at UCSD. Applicants are asked to submit 5 - 10 page proposals outlining a specific modelling project in a well-defined domain of developmental psychology. Criteria for evaluating proposals will include (1) the scientific merit and feasi- bility of the project itself (2) the applicant's computer sophistication and probability of success with short term training, (3) the probability that the applicant can and will continue working at the interface between neural model- ling and developmental psychology (including access to ade- quate computer facilities at the applicant's home site). Applicants should indicate the preferred duration and start- ing date for the training program. Applications should be submitted to Jeff Elman, Direc- tor, Center for Research on Language, University of California, San Diego, La Jolla, Ca. 92093. For further information, contact Jeff Elman (619-534-1147) or Elizabeth Bates (619-534-3007). Email inquiries may be sent to elman@amos.ling.ucsd.edu or bates@amos.ling.ucsd.edu. ------------------------------ Subject: No hidden neurons, BP vs Perceptrons. Report available. From: sontag@fermat.rutgers.edu Date: Fri, 02 Dec 88 10:45:58 -0500 The following technical report is now available from the Rutgers Center for Systems and Control. Please send requests to sycon@fermat.rutgers.edu including your complete mailing address. If an electronic version (latex file) is sufficient, please specify. (This is far better for us, since it saves printing and mailing costs.) -eduardo sontag ____________________________________________________________________________ Report SYCON-88-12 Backpropagation Separates when Perceptrons Do, E.D. Sontag and H.J. Sussmann, Nov. 88. (9 pages.) We consider in this paper the behavior of the least squares problem that arises when one attempts to train a feedforward net with no hidden neurons. It is assumed that the net has monotonic non-linear output units. Under the assumption that a training set is **separable**, that is that there is a set of achievable outputs for which the error is zero, we show that there are no non-global minima. More precisely, we assume that the error is of a **threshold** LMS type, in that the error function is zero for values "beyond" the target value. Our proof gives in addition the following stronger result: the continuous gradient adjustment procedure is such that **from any initial weight configuration** a separating set of weights is obtained **in finite time**. Thus we have a precise analogue of the perceptron learning theorem. We contrast our results with the more classical pattern recognition problem of threshold LMS with linear output units. ____________________________________________________________________________ NOTE: the report now includes comments about the relation with the works: Shrivastava, Y., and S. Dasgupta, ``Convergence issues in perceptron based adaptive neural network models,'' in {\it Proc.25th. Allerton Conf.Comm. Contr. and Comp.}, U.of Illinois, Urbana, Oct. 1987, pp. 1133-1141. and Wittner, B.S., and J.S. Denker, ``Strategies for teaching layered networks classification tasks,'' in {\it Proc. Conf. Neural Info. Proc. Systems,} Denver, 1987, Dana Anderson (Ed.), AIP Press. Both of these were brought to our attention after Geoff's posting to the net. In summary, the main difference with the latter is that our convergence theorem does allow for sigmoidal nonlinearities. But the idea that "thresholds" --or as Steve Hanson and others prefer, "margins," -- are needed was clearly stated in their paper, which should get all the credit in that regard. The main differences with the first of the above papers are also explained. ------------------------------ Subject: Stefan Shrier on Abduction Machines for Grammars From: pratt@zztop.rutgers.edu (Lorien Y. Pratt) Organization: Rutgers Univ., New Brunswick, N.J. Date: 02 Dec 88 21:54:04 +0000 This is the last talk of the semester. Thanks for helping to make this a successful colloquium series! --Lori Fall, 1988 Neural Networks Colloquium Series at Rutgers Abduction Machines for Grammar Discovery ---------------------------------------- Stefan Shrier Grumman-Ctec, McLean, VA Room 705 Hill center, Busch Campus Friday December 9, 1988 at 11:10 am Refreshments served before the talk Abstract Abduction machines (AMs) discover regularity structure in patterns. For language patterns (e.g., English sentences) several such machines demonstrate how they learn some aspects of language. The machines embody algorithms that train to learn word classes and grammars. These machines exhibit linguistic competence in the sense that they can produce and process "new" sentences to which they had not been exposed during training. A computer model, which simulates a learner, acquires an interesting subset of English grammar from another computer model which simulates a teacher who knows the language. Lorien Y. Pratt Computer Science Department pratt@paul.rutgers.edu Rutgers University Busch Campus (201) 932-4634 Piscataway, NJ 08854 ------------------------------ Subject: Stanford Adaptive Networks Colloquium From: netlist@psych.Stanford.EDU (Mark Gluck) Date: Mon, 05 Dec 88 06:57:35 -0800 Stanford University Interdisciplinary Colloquium Series: Adaptive Networks and their Applications Dec. 6th (Tuesday, 3:15pm) ************************************************************************** Self-Organization in a Perceptual Network RALPH LINSKER IBM T. J. Watson Research Center Yorktown Heights, New York Tel.: (914)-945-1077; e-mail: linsker@ibm.com ************************************************************************** Abstract What principles might help to account for the strikingly complex sets of feature-analyzing properties found in mammalian perceptual systems, and for their organization and integration? A Hebb-type synaptic modification rule causes model cells in a feedforward network to develop feature-analyzing proper- ties (R. Linsker, Proc. Natl. Acad. Sci. USA 83, 7508-12, 8390-94, 8779-83 (Oct.-Nov. 1986)). These include center- surround and orientation-selective cells (arranged in orien- tation columns) that have qualitative similarities to cells of the first several stages of the mammalian visual pathway. Furthermore, under certain conditions Hebb-type rules gener- ate model cells each of whose output activities conveys max- imum information about the input activity values presented to it (R. Linsker, Computer 21 (3) 105-117 (March 1988)). These results suggest a potential organizing principle, which I call "maximum information preservation," for each processing stage of a multilayered perceptual network having feedforward and lateral (intralayer) connections. According to this principle, each processing stage develops so that the output signal values (from that stage) jointly convey maximum information about the input values (to that stage), subject to certain constraints. The quantity that is maxi- mized is a Shannon information rate. I will discuss some consequences of the principle, and its possible role in bi- ological and machine perceptual systems. ************************************************************************** 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. Information: To be added to the network mailing list, netmail to netlist@psych.stanford.edu For additional information, contact Mark Gluck (gluck@psych.stanford.edu). Co-Sponsored by: Departments of Electrical Engineering (B. Widrow) and Psychology (D. Rumelhart, M. Pavel, M. Gluck), Stanford Univ. ------------------------------ Subject: TR from ICSI on "Knowledge-Intensive Recruitment Learning" From: baker%icsi.Berkeley.EDU@berkeley.edu (Paula Ann Baker) Date: Mon, 05 Dec 88 16:02:05 -0800 ********************************************************************* Technical Report available from the International Computer Science Institute "Knowledge-Intensive Recruitment Learning" TR-88-010 Joachim Diederich International Computer Science Institute 1947 Center Street Berkeley, CA 94704 Abstract The model described here is a knowledge-intensive connectionist learning system which uses a built-in knowledge representation module for inferencing, and this reasoning capability in turn is used for knowledge-intensive learning. The method requires only the presentation of a single example to build a new concept representation. On the connectionist network level, the central process is the recruitment of new units and the assembly of units to represent new conceptual information. Free, uncommitted subnetworks are connected to the built-in knowledge network during learning. The goal of knowledge-intensive connectionist learning is to improve the operationality of the knowledge representation: Mediated inferences, i.e. complex inferences which require several inference steps, are transformed into immediate inferences; in other words, recognition is based on the immediate excitation from features directly associated with a concept. This technical report is an extended version of: J. Diederich: Steps toward knowledge-intensive connectionist learning. To appear in: Pollack, J. & Barnden, J. (Eds.): Advances in Connectionist and Neural Computation Theory. Ablex Publ. 1988 Please send requests for copies by e-mail to: info@icsi.berkeley.edu or by post to: Librarian International Computer Science Institute 1947 Center Street, Suite 600 Berkeley, CA 94704 ************************************************************** ------------------------------ Subject: INTERFACE Call for Commentators and/or Original Contributions. From: MUSICO%BGERUG51.BITNET@CUNYVM.CUNY.EDU Date: Fri, 09 Dec 88 14:38:00 +0100 INTERFACE Call for Commentators and/or Original Contributions. -------------------------- MUSIC AND DYNAMIC SYSTEMS ========================= INTERFACE - Journal of New Music Research - is an international journal published by Swets & Zeitlinger B.V., Lisse, The Netherlands (this year vol. 17). It is devoted to the discussion of all questions which fall into the borderline areas between music on the one hand, physical and human sciences or related technologies on the other hand. New fields of research, as well as new methods of investigation in known fields receive special emphasis. INTERFACE is planning a special issue on MUSIC AND DYNAMIC SYSTEMS. The motivation comes from two sources : First there is the renewed interest in Dynamic Systems Theory from the point of view of massive parallel computing and artificial intelligence research. Massive parallel techniques and technology have very recently been applied to music perception/cognition and to strategies for automated composition. The approach is an alternative to the classical symbol-based approaches to cognition and problem solving and it is believed that it may establish a new paradigm that dominates research for the coming decennia. The second motivation comes from a recently received original contribution to INTERFACE by two Romenian scientists : Cosmin and Mario Georgescu. They propose a system approach to musicology based on the General Systems Theory. The paper ("A System Approach to Music") is challenging in that it raises a number of methodological problems (e.g. problems of verification) in musicology. The authors claim that "The paper should be considered primarily as an exposition of principles and as an argument in favour of the credibility degree of the system approach in musicology. The change of this approach into an effective analysis tool for musical work is a future task that goes beyond the aim of this paper.". However, General Systems Theory is by no means the only possible application of Systems Theory to music. The massive parallel approach in computing and the application of Dynamic Systems Theory to the field of music perception and cognition, automated compositional strategies, or historical musicology allows new insights in our understanding and comprehention of the complex phenomenon which we all admire. How far can we go in modeling the complex dynamics of MUSIC? -------------------------- - Contributions to this special issue of INTERFACE on MUSIC AND DYNAMIC SYSTEMS may be sent to Marc Leman before june 30 (publication of this issue is planned in the fall of 1989). - Commentators interested in the Georgescu's paper (61pp.) may ask for a copy. --------------------------- Please send your correspondence for this issue to : Marc Leman (editor) University of Ghent Institute for Psychoacoustics and Electronic Music Blandijnberg 2 B-9000 GHENT Belgium e-mail : musico@bgerug51.bitnet The address of the publisher is : Swets Publishing Service Heereweg 347 2161 CA Lisse The Netherlands ------------------------------ End of Neurons Digest *********************