neuron-request@HPLABS.HP.COM (Neuron-Digest Moderator Peter Marvit) (11/22/88)
Neuron Digest Monday, 21 Nov 1988 Volume 4 : Issue 27 Today's Topics: Neural Nets as A/D Converter Notes on Neural Networks (Two Experiments) Questionaire about Axon/Netset Neural Nets and Search Neural nets and edge detection Wanted: references to unsupervised learning Technical definition of trainability Stanford Adaptive Networks Colloquium Send submissions, questions, address maintenance and requests for old issues to "neuron-request@hplabs.hp.com" or "{any backbone,uunet}!hplabs!neuron-request" ------------------------------------------------------------ Subject: Neural Nets From: U89_APEREZ1@VAXC.STEVENS-TECH.EDU (Andres Perez) Date: Mon, 14 Nov 88 14:46:28 -0500 I am currently a Senior at Stevens Institute of Technology and I am working on my Senior Design Thesis on Neural Networks and their circuit implementation; more specifically, I'm using Neural Nets to build an A/D converter using Hopfield's Model as outlined in his article on IEEE trans. on circ. and sys., vol cas-33, no.5, may 1986, pp.533. Is there any information on the specific circuit implementation of Hopfield's model (such as type of Op.Amps., etc) or some documentation were this model is clearly explained; or maybe you could tell me where such information could be obtained. I appreciate the attention to this letter and thank you in advance. Andres E. Perez Mail: Stevens Tech P.O. Box S-1009 Hoboken, NJ 07030 (USA) BITNET-mail: U89_AEP@SITVXB [[ Editor's Note: I betray my ignorance about this particular subject. Perhaps some kind reader could help this fellow? -PM ]] ------------------------------ Subject: Notes on Neural Networks From: kanecki@VACS.UWP.WISC.EDU (David Kanecki) Organization: The Internet Date: 15 Nov 88 04:35:47 +0000 [[ Editor's Note: This was recently posted on AI-Digest. I have lightly edited it for format and the usual punctuation et al. Recently, Mr. Kanecki became a Neuron Digest subscriber as well. He is also looking for a position, but I would prefer not to publish resumes in the Digest. Contact him directly for a copy. -PM ]] Notes on Neural Networks: During the month of September while trying various experiments on neural networks I noted two observations: 1. Based on how the data for the A and B matrix are set up, the learning equation of: T w(n)=w(n-1)+nn(t(n)-o(n)*i (n) may take more presentations for the system to learn than A and B output. 2. Neural Networks are self correcting in that if an incorrect W matrix is given by using the presentation/ update process the W matrix will give the correct answers, but the value of the individual elements will differ when compared to a correct W matrix. Case 1: Different A and B matrix setup For example, in applying neural networks to the XOR problem I used the following A and B matrix: A H | H B - ------- |------ 0 0 0 | 0 0 0 1 0 | 0 1 1 0 0 | 0 1 0 1 1 | 1 1 My neural network learning system took 12 presentations to arrive at the correct B matrix when presented with the corresponding A matrix. The W matrix was: W(12) = | -0.5 0.75 | | -0.5 0.75 | | 3.5 -1.25 | For the second test I set the A and B matrix as follows: A H | B - ------------ 0 0 0 | 0 0 1 0 | 1 1 0 0 | 1 1 1 1 | 0 This setup took 8 presentations for my neural network learning system to arrive at a correct B matrix when presented with the corresponding A matrix. The final W matrix was: W(8) = | -0.5 -0.5 2.0 | Conclusion: These experiments indicate to me that a system's learning rate can be increased by presenting the least amount of extraneous data. -------------- Case 2: Self Correction of Neural Networks In this second experiment I found that neural networks exhibit great flexibility. This experiment turned out to be a happy accident. Before I had developed my neural network learning system I was doing neural network experiments by spreadsheet and hand transcription. During the transcription three elements in 6 X 5 W matrix had the wrong sign. For example, the resulting W matrix was: | 0.0 2.0 2.0 2.0 2.0 | |-2.0 0.0 4.0 0.0 0.0 | W(0)= | 0.0 2.0 -2.0 2.0 -2.0 | | 0.0 2.0 0.0 -2.0 2.0 | |-2.0 4.0 1.0 0.0 0.0 | | 2.0 -4.0 2.0 0.0 0.0 | W(24) = | 0.0 2.0 2.0 2.0 2.0 | |-1.53 1.18 1.18 -0.25 -0.15 | | 0.64 0.12 -0.69 1.16 -0.50 | | 0.27 -0.26 -0.06 -0.53 0.80 | |-1.09 1.62 0.79 -0.43 -0.25 | | 1.53 -1.18 -0.68 0.25 0.15 | By applying the learning algorithm it took 24 presentations the W matrix to give correct B matrix when presented with corresponding A matrix. But, when the experiment was run on my neural network learning system I had a W(0) matrix of: W(0) = | 0.0 2.0 2.0 2.0 2.0 | |-2.0 0.0 4.0 0.0 0.0 | | 0.0 2.0 -2.0 2.0 -2.0 | | 0.0 2.0 -2.0 -2.0 2.0 | |-2.0 4.0 0.0 0.0 0.0 | | 2.0 -4.0 0.0 0.0 0.0 | After 5 presentations the W(5) matrix came out to be: W(5) = | 0.0 2.0 2.0 2.0 2.0 | |-2.0 0.0 4.0 0.0 0.0 | | 0.0 2.0 -2.0 2.0 -2.0 | | 0.0 2.0 -2.0 -2.0 2.0 | | 2.0 -4.0 0.0 0.0 0.0 | Conclusion: Neural networks are self correcting but the final W matrix way have different values. Also, if a W matrix does not have to go through the test/update procedure the W matrix could be used both ways in that a A matrix generates the B matrix and a B matrix generates the A matrix as in the second example. ---------------- I am interested in communicating and discussing various aspects of neural networks. I can be contacted at: kanecki@vacs.uwp.wisc.edu or at: David Kanecki P.O. Box 93 Kenosha, WI 53140 ------------------------------ Subject: Questionaire From: cs162faw@sdcc18.ucsd.EDU (Phaedrus) Organization: University of California, San Diego Date: 15 Nov 88 18:14:44 +0000 About two weeks ago, I posted a desire for Axon/Netset information, I'm afraid my scope was much to small, considering I only received two responses. I'm sorry to muck up the newsgroup, but I really do need this information, and my posting disappeared after a week or so. If you've ever used a neural network simulator or if you have good opinions regarding representations. Provided is a questionnare regarding Neural-Networking/PDP. Information from this questionnare will be used to design a user interface for an industrial neural network program which may perform any of the traditional PDP problems (e.g., back prop, counter prop, constraint satisfact, etc). The program can handle connections set up in any fashion (e.g., fully connected, feed-back connected, whatever), and it can also toggle between syncronous or asyncronous modes. What we're really interested in is what you feel is "hard" or "easy" about neural net representations. 1. What type of research have you done ? 2. What type of research are you likely to do in the future ? 3. What is your programming background ? 4. What simulators have you used before ? What did you like about their interfaces ? 5. Have you used graphical interfaces before ? Did you like them ? Do you think that you could use them for research-oriented problems ? Why or why not ? 6. Do you prefer to work with numerical representations of networks ? Weight matrices ? Connection Matrices ? Why or why not ? 7. Would you like to use a graphical PDP interface if it could craft complicated networks easily ? Why or why not ? 8. Do you forsee any difficulties you might have with graphical interfaces ? Any other comments along the same vein will be appreciated. Your opinion is REALLY wanted, so please take 5 minutes and hit 'r-'!!! Thank you, James Shaw [[ Editor's note: I almost hesitated in including this, since the commercial overtones are a bit much. I personally would suggest that, rather than blindly polling the ubiquitous net, Mr. Shaw should either recruit some volunteers to do bone fide user-interface research or expend some energy and identify some selcted customers qua users. On the other hand, his questionnaire provides some interesting items to think about when considering ANN simulators. -PM ]] ------------------------------ Subject: Neural Nets and Search From: zeiden@ai.cs.wisc.edu (Matthew Zeidenberg) Date: 15 Nov 88 23:44:51 +0000 Does anyone know of any papers on neural network solutions to problems involving heuristic search? I do not mean optimization problems such as Traveling Salesman, although these are obviously related. ------------------------------ Subject: Neural nets and edge detection From: sda@cs.exeter.ac.uk (Steven Dakin) Date: 16 Nov 88 10:59:28 +0000 Recently I read an interesting article on applications of neural nets in the field of visual edge detection. Basically it showed that single layer nets couldn't be trained using a simple learning rule (eg. delta) to recognize lines of various orientations. The problem is that a set of weights to recognize verticals destroy all the information about horizontality inherent in the pattern of connectivity. The author proposed a solution using hidden units. So far so good. However some months on, and I can't for the life of me remember the reference or the author. I have a feeling it may be unpublished, but if anyone has come across this piece, or indeed any related work in the field, please mail me. I'll compile and mail a list if there's enough. Thanks. Steven Dakin (sda@uk.ac.exeter.cs) ------------------------------ Subject: Wanted: references to unsupervised learning From: Jose A Ambros-Ingerson (Dept of ICS, UC Irvine) <jose%ci7.ics.uci.edu@PARIS.ICS.UCI.EDU> Date: Wed, 16 Nov 88 18:29:16 -0800 I'm currently writing a survey paper on the subject of Unsupervised Learning. I intend to cover both traditional pattern classification techniques as well as neural network approaches. I will however, try to make an emphasis on biologically plausible models. I would greatly appreciate any suggestions as to reading material. In order to reduce duplications I include my present selections: Gail A. Carpenter and Stephen Grossberg, Neural Dynamics of Category Learning and Recognition: Structural Invariants, Reinforcement and Evoked Potentials, In Pattern Recognition and Concepts in Animals, People and Machines, M. L. Commons, S. M. Kosslyn, E. R. J. Herrnstein (eds), Lawrence Erlbaum Associates, 1986. Richard O. Duda and Peter E. Hart, Pattern Classification and Scene Analysis, John Wiley \& Sons, 1973. Gerald M. Edelman, George N. {Reeke, Jr}, Selective Networks Capable of Representative Transformations, Limited Generalizations, and Associative Memory, Proc. Natl. Acad. Sci., 79:2091-2095, 1982. Brian Everitt, Cluster Analysis, Second edition, 1980, Halsted Press. Leif H. Finkel, Gerald M. Edelman, Interaction of Synaptic Modification Rules within Populations, Proc.Natl.Acad.Sci.82:1291-1295,1985. Douglas H. Fisher, Knowledge Acquisition Via Incremental Conceptual Clustering, Machine Learning,2:139-172, 1987. K. Fukushima, Cognitron: {A} Self-organizing Multilayered Neural Network, Biol. Cybernetics, 20:121-136, 1975. A K. Fukushima, NeoCognitron: {A} Self-organizing Neural Network Model for a Mechanism of Pattern Recognition Unaffected by Shift in Position, Biol. Cybernetics, 36:193-202, 1980. Mark A. Gluck, James E. Corter, Information, Uncertainty, and the Utility of Categories, Proceedings of the Cognitive Science Society Conference, 1985, Lawrence Erlbaum. S. Grossberg, Adaptive Pattern Classification and Universal Recoding: {I}. Parallel Development and Coding of Neural Feature Detectors. Biol. Cybernetics, 23:121-134, 1976 S. Grossberg, Adaptive Pattern Classification and Universal Recoding: {II}. Feedback, Expectation, Olfaction, Illusions, Biol. Cybernetics, 23:187-202,1976 Teuvo Kohonen, Self-{O}rganization and Associative Memory, 1984, Springer-Verlag. Chr. {von der} Malsburg, Self-{O}rganization of Orientation Sensitive Cells in the Striate Cortex, Kibernetik, 14:85-100,1973. J. Pearson, L. Finkel, G. Edelman, Plasticity in the Organization of Adult Cerebral Cortical Maps: {A} Computer Simulation Based on Neuronal Group Selection ,The Journal of Neuroscience, 7:12:4209-4223,1987 David E. Rumelhart, David Zipser, Feature Discovery by Competitive Learning Cognitive Science, 1985, 9:75-112 T. Y. Young, T. W. Calbert, Classification Estimation and Pattern Recognition, 1974, Elsevier. Thanks in advance, Jose A. Ambros-Ingerson ArpaNet: jambros@ics.uci.edu Dept. of Information and Computer Science Phone: (714) 856-7310 University of California Irvine CA, 92717 ------------------------------ Subject: Technical definition of trainability From: jwang@cwsys2.CWRU.EDU (Jun Wang) Date: Wed, 16 Nov 88 22:06:05 -0500 I am currently working on research of theory and methodology of artificial neural net in a general setting from system point of view. My approach to the problem is by formalization, categorization and characterization. I hope a complete theory and methodology on neural system as a means of deriving decision rules can be developed based on in-depth analysis and synthesis. I got some elementary results in this direction. I am very interested in trainability of neural nets. The following is my definition of trainability from a working paper under preparation. It is a Tex file (I made some modification), I hope it is readable. ********************************************************************** Definition 3.10 (Trainability): Given architecture and propagation rule, and learning rule, an artificial neural net (ANN) is trainable if and only if a set of definite parameters $w$ can be obtained which result in minimum of errors, precisely, an ANN is trainable iff \forall \epsilon > 0, \exists T>0, \exists w(T)\in W, if t>= T || w(t+\delta t) - w(t)|| <= \epsilon and E=\min\sum^P_{p=1}\mu_p|| z^p- y(x^p, w(t)) ||_p where (x^p,z^p) \in S_{tr}. An ANN is globally trainable if it is trainable under arbitrary initial conditions. An ANN is globally and absolutely trainable if it is globally trainable at optimum parameters with respect to given E(w), i.e. \min_{w\in W} E(w(t))=\sum_{p=1}^P ||t^p - o^p(w(t))||_p=0, or \lim_{t\to\infty}E(w(t))=0. ************************************************************************* If anybody has some comments or suggestions on this property of neural nets, or knows someone has been working on this, please tell me via E-mail me or postal mail. Thanks. Jun Wang Dept. of Systems Engg. Case Western Reserve Univ. Cleveland, Ohio 44106 jwang@cwsys2.cwru.edu jwang@cwcais.cwru.edu ------------------------------ Subject: Stanford Adaptive Networks Colloquium From: netlist@psych.Stanford.EDU (Mark Gluck) Date: Thu, 17 Nov 88 17:10:19 -0800 Stanford University Interdisciplinary Colloquium Series: Adaptive Networks and their Applications Nov. 22nd (Tuesday, 3:15pm) ************************************************************************** Toward a model of speech acquisition: Supervised learning and systems with excess degrees of freedom MICHAEL JORDAN E10-034C Department of Brain and Cognitive Sciences Massachussetts Institute of Technology Cambridge, MA 02139 <jordan@psyche.mit.edu> ************************************************************************** Abstract The acquisition of speech production is an interesting domain for the development of connectionist learning methods. In this talk, I will focus on a particular component of the speech learning problem, namely, that of finding an inverse of the function that relates articulatory events to perceptual events. A problem for the learning of such an inverse is that the forward function is many-to-one and nonlinear. That is, there are many possible target vectors corresponding to each perceptual input, but the average target is not in general a solution. I will argue that this problem is best resolved if targets are specified implicitly with sets of constraints, rather than as particular vectors (as in direct inverse system identification). Two classes of constraints are distinguished---paradigmatic constraints, which implicitly specify inverse images in articulatory space, and syntagmatic constraints, which define relationships between outputs produced at different points in time. (The latter include smoothness constraints on articulatory representations, and distinctiveness constraints on perceptual representations). I will discuss how the interactions between these classes of constraints may account for two kinds of variability in speech: coarticulation and historical change. ************************************************************************** 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). Upcoming talks: Dec. 6: Ralph Linsker (IBM) Co-Sponsored by: Departments of Electrical Engineering (B. Widrow) and Psychology (D. Rumelhart, M. Pavel, M. Gluck), Stanford Univ. ------------------------------ End of Neurons Digest *********************