neuron-request@HPLABS.HP.COM ("Neuron-Digest Moderator Peter Marvit") (08/05/89)
Neuron Digest Friday, 4 Aug 1989 Volume 5 : Issue 33 Today's Topics: TR available: Optimum Supervised Learning Connection Science Journal CRG-TR-89-3 TR available: Contribution Analysis wanted: guest researcher Machine Learning Mailing List. Neural Network Session at SME Conference AUTOFACT Postdoc position at Bellcore Preprint Available: Configural-cue Network Model of Animal& Human Learning Ph.D. thesis available call for papers 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: TR available: Optimum Supervised Learning From: Manoel Fernando Tenorio <tenorio@ee.ecn.purdue.edu> Date: Tue, 06 Jun 89 17:05:00 -0500 The Tech Report below will be available by June, 15. Please do not reply to this posting. Send all you requests to jld@ee.ecn.purdue.edu Self Organizing Neural Network for Optimum Supervised Learning Manoel Fernando Tenorio Wei-Tsih Lee School of Electrical Engineering School of Electrical Engineering Purdue University Purdue University W. Lafayette, IN. 47907 W. Lafayette, IN. 47907 tenorio@ee.ecn.purdue.edu lwt@ed.ecn.purdue.edu Summary Current neural network algorithms can be classified by the following characteristics: the architecture of the network, the error criteria used, the neuron transfer function, and the algorithm used during learning. For example: in the case of back propagation, one would classify the algorithm as a fixed architecture (feedforward in most cases), using a MSE criteria, and a sigmoid function on a weighted sum of the input, with the Generalized Delta Rule performing a gradient descent in the weight space. This characterization is important in order to assess the power of such algorithms from a modeling viewpoint. The expressive power of a network is intimately related with these four features. In this paper, we will discuss a neural network algorithm with noticeably different characteristics from current networks. The Self Organizing Neural Network (SONN) [TeLe88] is an algorithm that through a search process creates the network necessary and optimum in the sense of performance and complexity. SONN can be classified as follows. The network architecture is constructed through a search using Simulated Annealing (SA),and it is optimum in that sense. The error criteria used is a modification of the Minimum Description Length Criteria called the Structure Estimation Criteria (SEC); it takes into account both the performance of the algorithm and the complexity of the structure generated. The neuron transfer function is individually chosen from a pool of functions, and the weights are adjusted during the neuron creation. This function pool can be selected with a priori knowledge of the problem, or simply use a class of non-linearities shown to be general enough for a wide variety of problems. Although the algorithm is stochastic in nature (SA), we show that its performance is extremely high both in comparative and absolute terms. In [TeLe88], we have used SONN as an algorithm to identify and predict chaotic series, particularly the Mackey-Glass equation [LaFa87, Mood88] was used. For comparison, the experiments of using Back Propagation for this problem were replicated under the same computational environment. The results indicated that for about 10% of the computational effort, the SONN delivered a 2.11 times better model (normalized RMSE). Some inherited aspects of the algorithm are even more interesting: there were 3.75 times less weights, 15 times less connections, 6.51 times less epochs over the data set, and only 1/5 of the data was fed to the algorithm. Furthermore, the algorithm generates a symbolic representation of the network which can be used to substitute it, or be used for the analysis of the problem. ****************************************************************************** We have further developed the algorithm, and although not part of the report above, it will be part of a paper submitted to NIPS'89. There, some major improvements on the algorithm are reported. The same chaotic series problem can now run with 26.4 less epochs over the data set that BP, and have generated the same model in about 18.5 seconds of computer time. (This is down from 2 CPU hours in a Gould NP1 Powernode 9080). Performance on a Sun 3-60 was sightly over 1 minute. These performance figures include the use of an 8 times larger function pool; the final performance now independs of the size of the pool. Other aspects of the algorithm are also important considering. Because of its stochastic nature, no two runs of the algorithm should be the same. This can become a hindrance if a suboptimal solution is desired, since at every run the set of suboptimal models can be different. A report on modifications of the original SONN to run on an A* search are presented. Since the algorithm generates partial structures at each iteration, the learning process is only optimized for the structure presently generated. If such substructure is used as a part of a larger structure, then no provision is made to readjust its weights making the final model slightly stiff. A provision for melting the structure (parametric readjustment) is also discussed. Finally, the combination of symbolic processing with this numerical method can lead to construction of AI-NN based methods for supervised and unsupervised learning. The ability of SONN to take symbolic constraints and produce symbolic information can make such a system possible. Implications of this design are also explored. [LaFa87] - Alans Lapedes and Robert Farber, How Neural Networks Work, TR LA-UR-88-418, Los Alamos, 1987. [Mood88] - J. Moody, Fast Learning in Multi-Resolution Hierarchies, Advances in Neural Information Processing Systems, D. Touresky, Ed., Morgan Kaufmann, 1989 (NIPS88). [TeLe88] - M. F. Tenorio and W-T Lee, Self Organizing Neural Networks for the Identification Problem, Advances in Neural Information Processing Systems, D. Touresky, Ed., Morgan Kaufmann, 1989 (NIPS88). ------------------------------ Subject: Connection Science Journal From: Lyn Shackleton <lyn@CS.EXETER.AC.UK> Date: Wed, 07 Jun 89 13:36:25 -0000 ANNOUNCEMENT Issue 1. of the new journal CONNECTION SCIENCE has just gone to press and Issue 2. will follow shortly. The editors are very pleased with the response they have received and would welcome more high quality submissions or theoretical notes. VOLUME 1 NUMBER 1 CONTENTS Michael C Mozer & Paul Smolensky 'Using Relevance to Reduce Network Size Automatically' James Hendler 'The Design and Implementation of Symbolic Marker-Passing Systems' Eduardo R Caianello, Patrik E Eklund & Aldo G S Ventre 'Implementations of the C-Calculus' Charles P Dolan & Paul Smolensky 'Tensor Product Production System: A Modular Architecture and Representation' Christopher J Thornton 'Learning Mechanisms which Construct Neighbourhood Representations' Ronald J Williams & David Zipser 'Experimental Analysis of the Real-Time Recurrent Learning Algorithm' Editor: Dr NOEL E SHARKEY, Centre for Connection Science, Dept of Computer Science, University of Exeter, UK Associate Editors: Andy CLARK (University of Sussex, Brighton, UK) Gary COTTRELL (University of California, San Diego, USA) James A HENDLER (University of Maryland, USA) Ronan REILLY (St Patrick's College, Dublin, Ireland) Richard SUTTON (GTE Laboratories, Waltham, MA, USA) FORTHCOMING IN VOLUMES 1 & 2 Special Issue on Natural Language, edited by Ronan Reilly & Noel Sharkey Special Issue on Hybrid Symbolic/Connectionist Systems, edited by James Hendler For further details please contact. lyn shackleton (assistant editor) Centre for Connection Science JANET: lyn@uk.ac.exeter.cs Dept. Computer Science University of Exeter UUCP: !ukc!expya!lyn Exeter EX4 4PT Devon BITNET: lyn@cs.exeter.ac.uk.UKACRL U.K. ------------------------------ Subject: CRG-TR-89-3 From: Carol Plathan <carol@ai.toronto.edu> Date: Fri, 16 Jun 89 15:33:15 -0400 The Technical Report CRG-TR-89-3 by Hinton and Shallice (May 1989) can be obtained by sending me your full mailing address. An abstract of this Report follows: LESIONING A CONNECTIONIST NETWORK: INVESTIGATIONS OF ACQUIRED DYSLEXIA ----------------------------------------------------------------------- Geoffrey E. Hinton Tim Shallice Department of Computer Science MRC Applied Psychology Unit University of Toronto Cambridge, UK ABSTRACT: - -------- A connectionist network which had been trained to map orthographic representation into semantic ones was systematically 'lesioned'. Wherever it was lesioned it produced more Visual, Semantic, and Mixed visual and semantic errors than would be expected by chance. With more severe lesions it showed relatively spared categorical discrimination when item identification was not possible. Both phenomena are qualitatively similar to those observed in neurological patients. The error pattern is that characteristically found in deep dyslexia. The spared categorical discrimination is observed in semantic access dyslexia and also in a form of pure alexia. It is concluded that the lesioning of connectionist networks may produce phenomena which mimic non-transparent aspects of the behaviour of neurological patients. ------------------------------ Subject: TR available: Contribution Analysis From: Dennis Sanger <sanger@boulder.Colorado.EDU> Date: Thu, 22 Jun 89 09:56:59 -0600 University of Colorado at Boulder Technical Report CU-CS-435-89 is now available: Contribution Analysis: A Technique for Assigning Responsibilities to Hidden Units in Connectionist Networks Dennis Sanger AT&T Bell Laboratories and the University of Colorado at Boulder ABSTRACT: Contributions, the products of hidden unit activations and weights, are presented as a valuable tool for investigating the inner workings of neural nets. Using a scaled-down version of NETtalk, a fully automated method for summarizing in a compact form both local and distributed hidden-unit responsibilities is demonstrated. Contributions are shown to be more useful for ascertaining hidden-unit responsibilities than either weights or hidden-unit activations. Among the results yielded by contribution analysis: for the example net, redundant output units are handled by identical patterns of hidden units, and the amount of responsibility a hidden unit takes on is inversely proportional to the number of hidden units. Please send requests to conn_tech_report@boulder.colorado.edu. ------------------------------ Subject: wanted: guest researcher From: unido!gmdzi!zsv!joerg@uunet.UU.NET (Joerg Kindermann) Date: Fri, 23 Jun 89 16:20:55 +0200 If you are a postgraduate student of scientist with a strong background in neural networks, we are interested to get in touch with you: We are a small team (5 scientists plus students) doing research in neural networks here at the GMD. Currently we are applying to get funding for several positions of guest researchers. But: we need strong arguments (i.e. good people who are interested in a stay) to actually get the money. Our research interests are both theoretical and application oriented. The main focus is on temporal computation (time series analysis) by neural networks. We are using multi-layer recurrent networks and gradient learning algorithms (backpropagation, reinforcement). Applications are speech recognition, analysis of medical data (ECG, ...), and navigation tasks for autonomous vehicles (2-D simulation only). A second research direction is the optimization of neural networks by means of genetic algorithms. We are using both SUN3s and a parallel Computer (64 cpu transputer-based). So, if you are interested, please write a letter, indicating your background in neural networks and preferred dates for your stay. Dr. Joerg Kindermann Gesellschaft fuer Mathematik und Datenverarbeitung mbH (GMD) Postfach 1240 email: joerg@gmdzi.uucp D-5205 St. Augustin 1, FRG phone: (+49 02241) 142437 ------------------------------ Subject: Machine Learning Mailing List. From: Michael Pazzani <pazzani@ICS.UCI.EDU> Date: Fri, 21 Jul 89 11:49:40 -0700 Could you include this message in the next neuron mailing? Machine Learning Mailing List A new mailing list, ML-LIST is being formed. ML-LIST will be devoted to discussions of issues concerning machine learning and announcements of interest to the machine learning community, including neural network learning algorithms. The list is moderated by Michael Pazzani of University of California, Irvine. The motivation for the list is to promote informal discussions of research topics and methodology. The list will be mailed out approximately once a week. To reduce the volume of information and to maintain a forum for meaningful interaction, many topics not appropriate to the scientific treatment of computational approaches to learning will be avoided. These include: * Questions on programming languages * Extended philosophical debates * Flames * Why are all the machine learning faculty at Irvine so short? * Messages with many spelling and grammar errors Messages will be encouraged on topics such as: * Discussions of research results * Discussions of research methodology * Brief announcements of the availability of Tech Reports * Abbreviated Calls for Papers & Conference Announcements In order to avoid unnecessary mail traffic, local redistribution of the ML-LIST, is encouraged. To subscribe to ML-LIST send mail including an e-mail address to: ml-request@ics.uci.edu Before rushing to subscribe, talk to your System Administrator about your local facilities for mail redistribution. Submissions to ML-LIST are e mailed to: ml@ics.uci.edu Thanks Michael Pazzani ------------------------------ Subject: Neural Network Session at SME Conference AUTOFACT From: David Kanecki <kanecki@vacs.uwp.wisc.edu> Date: Tue, 25 Jul 89 16:46:18 -0500 I received a flyer from the Society of Manufacturing Engineers(SME) about the AUTOFACT conference being held in Detroit, Michigan on October 30th thru November 2nd. Session 29 on Neural Networks includes talks as "Neural Network Basics", "An Artificial Neural System for Reading Printed Credit Card Numbers", "Optical Neural Networks for Process Control" and "Neurocomputing: Applications for an Important New Data Processing Technology" The theme for the conference is partnership for integration which in my last paper to the neuron digest was the basis for the paper I sent. To receive more information contact: AUTOFACT Conferene SME Conference Regiostrar Technical Activities One SME Drive P.O. Box 930 Dearborn, MI 48121-0930 Phone: (313)-271-1080 Fax: (313)-271-2861 ------------------------------ Subject: Postdoc position at Bellcore From: Joshua Alspector <josh@flash.bellcore.com> Date: Fri, 28 Jul 89 16:41:32 -0400 POSTDOCTORAL POSITION IN NEURAL NET RESEARCH The Adaptive Systems Research Group at Bellcore, Morristown, NJ is looking for a postdoctoral researcher for a period of 1 - 2 years starting approximately November, 1989. Bellcore has a stimulating research environment for neural computation with active programs in neural network theory and analysis, in applications such as speech recognition and expert systems, and in optical and electronic implementation. This is an excellent opportunity for a researcher to be exposed to this environment while contributing to the effort in VLSI implementation. A test chip that implements a neural network based learning algorithm related to the Boltzmann machine has been designed, fabricated and tested. The next step is to implement a useful, multiple-chip system that can learn to solve difficult artificial intelligence problems. We will extend our study of electronic implementation issues to large scale systems using a three-pronged approach: 1) Further development of learning algorithms and architectures suitable for modular VLSI implementation. To be useful, algorithms must be implementable because learning by example takes too long using serial computer simulations. Therefore, the algorithms should take into account the constraints imposed by VLSI. 2) Functional simulation of large scale hardware systems using benchmark test problems. We will build a computer-based development system for testing algorithms in software. This will be composed of software modules, some of which eventually will be replaced by hardware learning modules. A computation module may be run on a remote parallel machine. This will serve as a platform for algorithm development, will perform functional simulation of a hardware system before design, and also will be the front end for testing the chips and boards after fabrication. 3) Design and fabrication of prototype chips suitable for inclusion in such systems. As a first step in the development of large scale, modular VLSI systems, our learning test chip will be expanded to contain more neurons and synapses and to enable construction of a multichip system. This system would be taken to board-level design and fabrication. Evaluation will involve a speed comparison using a variety of benchmarks of three neural network implementations: software on a serial machine, software on a general purpose parallel machine, and special purpose neural hardware using the board level system we build. Chip and board design will be carried out using a combination of sophisticated VLSI CAD tools. The successful candidate should be involved in many aspects of this work including the design of algorithms and architectures for VLSI neural implementation, computer programming to simulate and test the existing and proposed neural architectures, and the design of analog and digital chips to implement them. He or she should be capable of doing independent publishable research in neural network learning theory, in parallel software simulation, in applications of neural information processing, or in VLSI implementations of neural network learning models. Please enclose a resume, a copy of a recent paper, and the names, addresses, and phone numbers of three references. Send applications to: Joshua Alspector Bellcore, MRE 2E-378 445 South St. Morristown, NJ 07960-1910 ------------------------------ Subject: Preprint Available: Configural-cue Network Model of Animal& Human Learning From: gluck@psych.Stanford.EDU (Mark Gluck) Date: Mon, 31 Jul 89 09:59:01 -0700 Pre-print of short conference paper available: Eleventh Annual Conference of the Cognitive Science Society, Ann Arbor, MI. August 16-19, 1989. A CONFIGURAL-CUE NETWORK MODEL OF ANIMAL AND HUMAN ASSOCIATIVE LEARNING Mark A. Gluck Gordon H. Bower Michael R. Hee Dept. of Psychology Stanford University Jordan Hall; Bldg. 420 Stanford, CA 94305 (415) 725-2434 Email: gluck@psych.stanford.edu ABSTRACT We test a configural-cue network model of human classification and recognition learning based on Rescorla & Wagner's (1972) model of classical conditioning. The model extends the stimulus representation assumptions from our earlier one-layer network model (Gluck & Bower, 1988b) to include pair-wise conjunctions of features as unique cues. Like the exemplar context model of Medin & Schaffer (1978), the representational assumptions of the configural-cue network model embody an implicit exponential decay relationship between stimulus similarity and and psychological (Hamming) distance, a relationship which has received substan- tial independent empirical and theoretical support (Shepard, 1957, 1987). In addition to results from animal learning, the model accounts for several aspects of complex human category learning, including the relationship between category similarity and linear separability in determining classification difficulty (Medin & Schwanenflugel, 1981), the relationship between classif- ication and recognition memory for instances (Hayes-Roth & Hayes-Roth, 1977), and the impact of correlated attributes on classification (Medin, Altom, Edelson, & Freko, 1982). ------------------------------ Subject: Ph.D. thesis available From: swain@cs.rochester.edu Date: Tue, 01 Aug 89 12:03:00 -0400 [[ Editor's Note: Please see the end of this annoucnement that this publication costs $7.50! -PM]] The following Ph.D. thesis now available: PARALLEL OBJECT RECOGNITION FROM STRUCTURE (THE TINKERTOY PROJECT) Paul R. Cooper Department of Computer Science University of Rochester Technical Report 301 July 1989 Abstract: This thesis examines the problem of recognizing structurally composed objects. The task is the recognition of Tinkertoys --- objects whose identity is defined solely by the spatial relationships between simple parts. Ultimately, a massively parallel framework incorporating a principled treatment of uncertainty and domain dependence is developed to address the problem. The basic architecture of the solution is formed by posing structure matching as a part-wise correspondence problem in a labelling framework, then applying the unit/value principle. The solution is developed incrementally. Complexity and correctness analyses and implementation experiments are provided at each phase. In the first phase, a special purpose network implementing discrete connectionist relaxation is used to topologically discriminate between objects. In the second step, the algorithm is generalized to a massively parallel formulation of constraint satisfaction, yielding an arc consistency algorithm with the fastest known time complexity. At this stage the formulation of the application problem is also generalized, so geometric discrimination can be achieved. Developing an implementation required defining a method for the domain specific optimization of the parallel arc consistency algorithm. The optimization method is applicable to arbitrary domains. In the final phase, the solution is generalized to handle uncertain input information and statistical domain dependence. Segmentation and recognition are computed simultaneously by a coupled Markov Random Field. Both problems are posed as labelling problems within a unified high-level MRF architecture. In the segmentation subnet, evidence from the image is combined with clique potentials expressing both qualitative {\em a priori} constraints and learnable domain dependent knowledge. Matching constraints and coupling constraints complete the definition of the field. The effectiveness of the framework is demonstrated in experiments involving the traditionally difficult problems of occlusion and accidental alignment. ============ TO ORDER, send requests to tr@cs.rochester.edu or physical mail to: Technical Reports Librarian, Department of Computer Science, University of Rochester, Rochester, NY 14627. The cost is $7.25. Make checks payable to the University of Rochester. ------------------------------ Subject: call for papers From: <OOMIDVAR%UDCVAX.BITNET@CORNELLC.cit.cornell.edu> Date: Fri, 04 Aug 89 11:18:00 -0400 Would you please post the following call for papers: Call For Papers International Journal of Computer Aided VLSI Design The above journal will publish a special issue on Neural Networks. Authors are invited to submit original manuscripts describing the recent advances in modelling, design, development, and VLSI Implementation of neural networks. Other suggested topics are neurocomputers, and their applications in the areas of vision, speech, self organization and control. Submit five copies of full-length paper fifteen to twenty pages long by the end of August 1989 to Dr. Omid M. Omidvar, Department of Computer Science, University of the District of Columbia, 4200 Connecticut Avenue, N.W., Washington, D.C. 20008. Phone (202) 282-7345. Fax: (202) 282-3677. email: oomidvar@udcvax.bitnet. ------------------------------ End of Neurons Digest *********************