neuron-request@HPLABS.HPL.HP.COM ("Neuron-Digest Moderator Peter Marvit") (02/03/90)
Neuron Digest Friday, 2 Feb 1990 Volume 6 : Issue 9 Today's Topics: Parallelism, Real vs. Simulated: A Query Network for coordinate transformations need help for Prof. Gallant's e-mail address NNs on Transputers Re: Data Complexity Our experience with NWorks by NeuralWare Emperor's New Mind: BBS Call for Commentators VLSI hardware for Artificial Neural Nets ghost in the hippocampus graph matching Bibliography (followup) ridiculous price Publishers of NN Journals Request For Info A "half-baked" Question... Re: Neuron Digest V6 #1 Signature verification request to neuron-digest readers address/bibliography Re: ND V6 #2 Help! Job Opening - Please Post 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: Parallelism, Real vs. Simulated: A Query From: harnad@clarity.Princeton.EDU (Stevan Harnad) Date: Thu, 05 Oct 89 23:36:54 -0400 I have a simple question: What capabilities of PDP systems do and do not depend on the net's actually being implemented in parallel, rather than just being serially simulated? Is it only speed and capacity parameters, or something more? Stevan Harnad Psychology Department Princeton University harnad@confidence.princeton.edu [[ Editor's note: Interesting question. I would suspect certain time-related aspects of biological parallelism may escape serial modeling, but I seem to recall general assertions atht *any* parallel porcess can be modelled sequentially. Of course, the reverse may be a bit tricky, as I'm discovering in my current literatire search for connectionist models of serial processing. Readers, any comments? -PM ]] ------------------------------ Subject: Network for coordinate transformations From: RM5I%DLRVM.BITNET@CUNYVM.CUNY.EDU Date: Mon, 23 Oct 89 15:35:16 -0500 Hi, I'm trying to implement the cartesian/polar and cartesian/sphere coordinate transformation in a backpropagation network. I was successful with cartesian/polarcoordinate transformation by using a two layer backpropagation network in which the first hidden layer has a tanh transfer function and the second layer is split into two halfs of which one has a sin transfer function the the other half has a sigmoid transfer function. Is someone else interested in this and can give me soem ideas of the more complex cartesian/spherecoordinate transformation ? Regards..Roland Luettgens (rm5i@dlrvm) Bitnet ------------------------------ Subject: need help for Prof. Gallant's e-mail address From: HCFU%TWNCTU01.BITNET@CUNYVM.CUNY.EDU Date: Wed, 15 Nov 89 13:57:00 +0800 I am current reading a paper written by Dr. Stephen I. Gallant in the Feb. 1988 issue of CACM. He suppose to be with the college of computer science of Northeasten University. In additoin, I am also interested in reading and collecting papers on the Neural nets in connection with expert systems. Would you pls give some suggestions, such as name of proceedings, TRs, special issues or models, and real systems etc. Thanks in advances Hsin Chia Fu ------------------------------ Subject: NNs on Transputers From: M Norman <mgn%castle.edinburgh.ac.uk@NSFnet-Relay.AC.UK> Date: Fri, 01 Dec 89 20:02:06 +0700 Herve Frylander (hfry@alize.imag.fr) requests information about parallelisation of NNs. The Edinburgh Concurrent Supercomputer has some experience in this and indeed a BackProp simulator called Rhwydwaith which runs in parallel on transputers. A useful reference is Norman Radcliffe Richards Smieja Wallace Collins Hayward and Forrest "Neural Network applications in the Edinburgh Concurrent Supercomputer Project" Neuro Computing: Algorithms Applications and Architectures eds. F. Fogelman Soulie and J Herault, to appear (1989). Edinburgh Preprint No 89/462 The contact point for Rhwydwaith is Nick Radcliffe (njr@uk.ac.ed.castle). Mike Norman Dept of Physics University of Edinburgh ------------------------------ Subject: Re: Data Complexity From: chaos%gidrah.Berkeley.EDU@jade.berkeley.edu (Jim Crutchfield) Date: Mon, 04 Dec 89 13:28:05 -0800 An element of the set "nonlinear dynamics people" would like to draw your attention to the following papers that address nonlinear model reconstruction and complexity in dynamical systems. Many of the questions brought up in the recent postings concerning "Data Complexity" receive constructive and quantitative answers when addressed to modeling nonlinear time-dependent processes. o "Equations of Motion from a Data Series", JPC and B. McNamara, Complex Systems 1 (1987) 417. o "Inferring Statistical Complexity", JPC and K. Young, Physical Review Letters 63 (1989) 105. o "Computation at the Onset of Chaos", JPC and K. Young, in Entropy, Complexity, and the Physics of Information, W. Zurek, editor, Addison-Wesley, Reading, Massachusetts (1989) in press. o "Information and its Metric", JPC, in Nonlinear Structures in Physical Systems - Pattern Formation, Chaos, and Waves, L. Lam and H. C. Morris, editors, Springer-Verlag, Berlin (1990) in press. o "Inferring the Dynamic, Quantifying Physical Complexity", JPC, in Quantitative Measures of Dynamical Complexity in Nonlinear Systems, A. M. Albano, N. B. Abraham, P. E. Rapp, and A. Passamante, eds., Plenum Press, New York (1989) in press. Also, I would like to point out that in the context of dynamical systems theory it is a theorem that the Kolmogorov-Chaitin complexity, based on the computational model of deterministic Turing machines, of a typical orbit of a chaotic dynamical system is degenerate with the system's metric entropy. The latter is based on Shannon information theory as introduced into dynamics by Kolmogorov and Sinai. Thus, from the viewpoint of stationary dynamical systems the K-C complexity is not of much interest. One might as well use Shannon information. In the above work we define a different complexity based on Turing machines with a random register (Bernoulli-Turing machines). This leads to a quantitative measure of physical complexity that is complementary to information-based measures of the degree of randomness. This is a new invariant for dynamical processes. One result is that the space of dynamical systems appears to be organized into informational phases (gas, liquid, and solid) the boundaries of which support high levels of computation. Physicists call the boundaries phase transitions. This result appears to hold not only for continuous dynamical systems, but also for discretized dynamical systems such as cellular automata. Note that although the K-C complexity is in general noncomputable, requiring the minimal Turing machine representation of given data, in the case of chaotic dynamical systems it can be readily estimated! (See J. P. Crutchfield and N. H. Packard, "Symbolic Dynamics of One-Dimensional Maps: Entropies, Finite Precision, and Noise", International Journal of Theoretical Physics volume 6/7 (1982) 433.) So much for advertising. What does this have to do with neural networks? Good question. Jim Crutchfield Physics Department University of California Berkeley, California 94720 (415) 642-1287 ------------------------------ Subject: Our experience with NWorks by NeuralWare From: RM5I%DLRVM.BITNET@CUNYVM.CUNY.EDU Date: Mon, 04 Dec 89 17:16:33 -0500 We use a Neural Network simulator called NWorks made by NeuralWare. It has 13 models of neural nets included like BP and ART. You can either design your own net with a ADD LAYER and ADD PROCESSING ELEMENT feature or use one of the standard nets with INSTANET or LOAD NETWORK. This software can run on different machines like IBM-PC and i think SUN. It has a graphics interface and you can view the network at learning intervalls and see graphically how weights get adjusted. This software costs a few hundred dollar but it is worth to buy this if you can. From my point of view it is a very useful simulator with a powerful userinterface and environment. If you have questions feel free to call me. Roland Luettgens (rm5i@dlrvm) Bitnet ------------------------------ Subject: Emperor's New Mind: BBS Call for Commentators From: harnad@clarity.Princeton.EDU (Stevan Harnad) Date: Tue, 05 Dec 89 01:01:46 -0500 [[Editor's Note: For those of you with access to USENET, there was quite a debate raging over Penfield's book, as well as the recent Scientific American articles by Searle and the Churchlands. i don't know if any of the mailing lists are carrying the discussions. I highly recommend the Scientific American articles, especially since the Churchlands are succint about viability of connectionist models with respect to their capabilities and potentail for "thinking." -PM ]] Below is the synopsis of a book that will be accorded a multiple book review (20 - 30 multidisciplinary reviews, followed by the author's response) in Behavioral and Brain Sciences (BBS), an international, interdisciplinary journal that provides Open Peer Commentary on important and controversial current research in the biobehavioral and cognitive sciences. Reviewers must be current BBS Associates or nominated by a current BBS Associate. To be considered as a reviewer for this book, to suggest other appropriate reviewers, or for information about how to become a BBS Associate, please send email to: harnad@confidence.princeton.edu or write to: BBS, 20 Nassau Street, #240, Princeton NJ 08542 [tel: 609-921-7771] ____________________________________________________________________ THE EMPEROR'S NEW MIND: CONCERNING COMPUTERS, MINDS AND THE LAWS OF PHYSICS Roger Penrose Rouse Ball Professor of Mathematics University of Oxford The Emperor's New Mind is an attempt to put forward a scientific alternative to the viewpoint of "Strong AI," according to which mental activity is merely the acting out of some algorithmic procedure. John Searle and other thinkers have likewise argued that mere calculation does not, of itself, evoke conscious mental attributes, such as understanding or intentionality, but they are still prepared to accept that the action of the brain, like that of any other physical object, could in principle be simulated by a computer. In my book I go further than this and suggest that the outward manifestations of conscious mental activity cannot even be properly simulated by calculation. To support this view I use various arguments to show that the results of mathematical insight, in particular, do not seem to be obtained algorithmically. The main thrust of this work, however, is to present an overview of the present state of physical understanding and to show that an important gap exists at the point where quantum and classical physics meet, and to speculate on how the conscious brain might be taking advantage of whatever new physics is needed to fill this gap, in order to achieve its non-algorithmic effects. ------------------------------ Subject: VLSI hardware for Artificial Neural Nets From: Hsin Chia Fu <HCFU%TWNCTU01.BITNET@CUNYVM.CUNY.EDU> Date: Sat, 09 Dec 89 09:30:00 +0800 I am interested in this particular VLSI hardware implementaion for ANN. Would some one out there give list of individuals or organizations on this matter. I would be very appreciated if the list could include both postal and electronic addresses. In addition, someone may pass addresses for Dr. Federico Faggin and/or Synaptics, Inc. Thanx Hsin Chia Fu [[ Editor's Note: I immediately think of Carver Mead's recent book on Analog VLSI. He describes the articial retina and cochlea he's been working on for years; in biological systems, these structures are often seen simply as "brain extensions" rather than ennervated organs. Thus they are about as "neural net" as one can get. -PM ]] ------------------------------ Subject: ghost in the hippocampus From: arti6!tony@relay.EU.net (Tony Bell) Date: Mon, 11 Dec 89 01:25:59 +0100 Thank you, Barry Kort, for the interesting account of the talks you heard at MIT. I was particularly intrigued to hear that Terry Sejnowski has reported that: "The Hebbs Synapse would seem to be the foundation for superstitious learning." So Nancy Reagan can blame it all on those NMDA receptors. ------------------------------ Subject: graph matching From: roysam@ecse.rpi.edu (Roysam) Date: Tue, 12 Dec 89 15:49:22 -0500 I have a citation on this subject that I would like to see, but really don't know how. "Matching of Attributed and non-attributed graphs by use of the Boltzmann Machine algorithm," Kuner, Siemens, West Germany. This paper was on the list of abstracts for a recent conference. Badri (roysam@ecse.rpi.edu) ------------------------------ Subject: Bibliography (followup) From: chuck@utkux1.utk.edu (chuck) Date: Tue, 12 Dec 89 17:26:14 -0500 Friends, The flurry of requests for references prompts me to mention several resources of a general nature which may be of interest. I hope these will be useful. A large bibliography, assembled and maintained by Eugene Miya, is available by anonymous ftp from icarus.riacs.edu (128.102.16.8) in the /pub/bib directory. This is primarily a bibliography of parallel and supercomputing references, but does contain quite a bit of other 'connectionist' stuff. It is in 'refer' format. Eugene Miya has asked to maintain the list, and requests that additions and corrections be sent to him (instructions are in the files in the same directory, along with several tools to assist in moving the format to scribe or tex.) I hope, in the near future, to contribute references to this bibliography which will include announcements and lists appearing on the internet in various places --including the archives of Neuron Digest. These entries will be primarily 'connectionist' references. Our library reports the acquisition of a reference book entitled "The 1989 Neuro-Computing Bibliography" by Casimir Klim. I have not had the opportunity to examine it, however, and cannot offer any comment at this time. Season's Greetings! Chuck Joyce chuck@cs.utk.edu ------------------------------ Subject: ridiculous price From: "Rolf Pfeifer" <pfeifer@ifi.unizh.ch> Date: 17 Dec 89 12:07:00 +0100 I ordered the IJCNN-89 proceedings from a book store in Zurich, Switzerland. They charged me a ridiculuous price of SFr. 520 (which is approximately US $ 330). This is clearly not meant to enhance the process of scientific communication. I wonder who gets how much on this. --Rolf Pfeifer AI Lab, Computer Science Department University of Zurich Zurich, Switzerland ------------------------------ Subject: Publishers of NN Journals From: Song Y. Yan <munnari!ysy@uunet.UU.NET> Date: Sun, 07 Jan 90 00:48:41 +1100 Do you know who are the publishers of the following two NN journals 1. Journal of Neural Network Computing 2. Neural Computation and their addresses? Thank you very much. Looking forward to hearing from you soon. Fondest Regards, Song Y. Yan S. Y. Yan, Dept of Computer Science, University of Melbourne, Parkville, Victoria 3052, Australia Phone: 61 3 344-6807 E-mail: ysy@cs.mu.oz.au ------------------------------ Subject: Request For Info From: "DAVE MCKEE" <mckee@tisss.radc.af.mil> Date: 09 Jan 90 14:08:00 -0400 For readers of the Neural Network E-Mail List. I am an Engineer at Rome Air Development Center, Griffiss AFB, Rome NY. I am interested in collecting information on reliability issues in neural network designs. Below is a edited version of a RADC Weekly Activity Report issued throughout RADC for information references. If you are interested in, or have information pertaining to these issues please send E-mail to the below address: MCKEE@TISSS.RADC.AF.MIL Thank you very much. David T. McKee Neural Network Reliability Characterization: Rome Air Development Center (RADC) at Griffiss AFB, Rome, NY 13441-5700 have initiated a library search for reliability oriented work in this area, hoping to also identify potential customers and sponsors. There is increasing worldwide interest in modeling, software simulation (using conventional and modified architecture machines), and direct silicon implementation of neural network emulations. (IEEE MICRO, Dec. 1989). DARPA is mega-funding research, and Ford, GTE, and even the USAF are pursuing real world applications to product debug, lifetime estimation based on complex process control data, and field failure prediction based on experience data bases. At this point we need to be alert to reliability implications being made: extensive interconnect may dominate reliability of some implementations; some network chips with many processing defects still work fairly well, without repair; optoelectronics technology may be used; analog CMOS (with its potential ESD and latch-up weaknesses) may dominate; etc. We also need to be alert to how they might be useful to us. It seems clear that researchers hope that networks will eventually offer competitive solutions to complex control and optimization problems. (Perhaps related to IC design, processing, inspection and screening)? Please help us collect references. ------------------------------ Subject: A "half-baked" Question... From: "DAVE MCKEE" <mckee@tisss.radc.af.mil> Date: 09 Jan 90 15:29:00 -0400 I am inquiring, under the heading "half baked ideas" what work, if any, has been done concerning neural networks that take into account quantum effects, morphogenic fields, or random processes where "true" random noise is used (as opposed to pseudo random algorithms)? I have been thinking how the neurons in the brain are affected by these processes, and I wonder how much of biological neural behavior is in fact affected by them. [[ Editor's note: See also Penfield's book, mentioned above. In this context, though, I'm not sure about the practical distinction betwen "true" and "pseudo" randomness, given the inherent limitations of computational algorithms. -PM ]] ------------------------------ Subject: Re: Neuron Digest V6 #1 From: Daniel Abramovitch <danny@hpldya> Date: Wed, 10 Jan 90 07:59:22 -0800 > From: worden@ut-emx.UUCP (worden) > Organization: The University of Texas at Austin, Austin, Texas > Date: 21 Oct 89 08:06:41 +0000 > > Lyapunov functions and stability criteria are one of the mainstays of > control theory (aka, linear systems theory). Linear systems theory is a subset of control theory. Much of control theory deals with nonlinear systems. In particular, Lyapunov functions are useful because they allow the engineer to establish the stability of nonlinear systems. (For linear systems, there are quite a few other methods of establishing the stability of a system.) The best (most intuitive) discussion of Lyapunov functions that I have come accross is in a book by Ogata, "Modern Control Engineering", published by Prentice Hall. One chapter is devoted to Lyapunov functions (about 20 pages). Danny Abramovitch danny%hpldya@hplabs.hp.com ------------------------------ Subject: Signature verification From: rmurrays%computer-science.strathclyde.ac.uk@NSFnet-Relay.AC.UK Date: Fri, 12 Jan 90 11:26:35 +0000 Dear World, I'm doing some research into the verification of handwritten signatures. The hardware has already been built and this supplies the verification system with 36 parameters, describing features of the signature. From this hardware a large database of hundreds of valid signatures has been developed. I am hoping to use the back-propagation algorithm for the verification of the signatures. The main difficulty will be in deciding how to mark the boundary between the valid decision area and the invalid area, so that the algorithm has two classes to decide on. We have to create the samples for the second class, the invalid decision area. Has anyone else done any work in this type of area? Are there already many well-established traditional multivariate statistical methods that I am ignorant of? Are there more suitable algorithms than back-propagation? I also hope to implement a recognition system, which will take the description of the signature and perform a sort of hashing function on it, to return the name of the author from the database. Any ideas or experience with this problem will also be carefully studied. Any help on either of these topics will be gratefully received and passed on to anyone else interested. Roderick Murray-Smith ------------------------------ Subject: request to neuron-digest readers From: UNNI%RCSMPB@gmr.com Date: Fri, 12 Jan 90 14:43:00 -0500 I am involved in organizing a workshop on neural networks for graduate students in India. Does anybody have public domain software for simulation of various networks on garden variety IBM PCs? This would be used by a group of about 30 students. Will any of the commercial software developers be willing to donate a simulation package for this workshop? unni@gmr.com ------------------------------ Subject: address/bibliography From: levine@antares.mcs.anl.gov Date: Mon, 15 Jan 90 12:11:07 -0600 Hi, not sure what email address to use to ask this question, and the question may be dumb, but... In the recent bibliography posted. How is one to tell the magazine, journal,... name. E.g., I recognize as Scientific American, but if I hadn't ....? Thanks --dave levine %A A. K. Dewdney %D 1985 %T Computer Recreations: Exploring the field of genetic algorithms in a primor\dial computer sea full of flibs %P 21-32 David Levine levine@mcs.anl.gov Mathematics and Computer Science {alliant,sequent,rogue}!anlams!levine 9700 Cass Avenue South (708) 972-6735 Argonne National Laboratory (708)-972-5986 (FAX) Argonne, Illinois 60439 ------------------------------ Subject: Re: ND V6 #2 From: gt0228b@prism.gatech.edu (gt0228b FALCO,VINNIE) Date: Tue, 23 Jan 90 00:06:11 -0500 In reference to the first article in the aforementioned 'Digest, I am wondering what the importance of specification (8) in that all Symbol Systems are SEMANTICALLY INTERPRETABLE : Surely there are symbol systems that have no semantic interpretation, yet are self consistent. The constraint of semantic interpretations upon symbol systems seems to rule out a whole class of systems that may have importance, while not having an implicit, intrinsinc semantic interpretation. - Vinnie Falco gt0228b@prism.gatech.edu ------------------------------ Subject: Help! From: Sanjeev Sharma <sharma@hpihoed> Date: Fri, 26 Jan 90 17:26:31 -0800 Hello fellow netters, I have a pattern-recognition and classification problem for which I am trying to determine the "best" ANN algorithm and architecture to use. Being a relative new-comer to the Neural Network field, I would appreciate any suggestions from the ANN gurus out there. The problem I'm trying to solve is described below: The input data is a set of real-valued 2-dimensional patterns with unknown temporal and spatial probability distributions. The input to the neural network is an n-tuple obtained from sampling these patterns at n-locations. Typically, n = 45. There are m binary (ON/OFF) outputs, any combination of which can be ON at one time (each output represents a certain condition detected in the input). However, generally, at most 3 of the outputs will be ON at one time. Typically, m = 10. At time k, the output vector W(k) is a function of the present and past input vectors V(k), V(k-1), ..., V(k-p). For the sake of simplicity, let p = 5. The network will be trained in a supervised-learning environment, with the expected output known for each input training vector. After the training period, the network will be used to classify input patterns P. These patterns P are similar to the training set T to the extent that both P and T are the products of the same underlying process. Note that the network will be required to exhibit the properties of rotational and translational invariance. Any suggestions, pointers, related work, etc. will be highly appreciated. I hope I've been clear in the description of my problem. If not, please feel free to drop me a line. Thanks. Sanjeev Sharma sharma%hpda@hplabs.hp.com ------------------------------ Subject: Job Opening - Please Post From: "Fabio Idrobo" <psy9a3n@BUACCA.BU.EDU> Date: Sat, 27 Jan 90 10:22:41 -0500 JOB ANNOUNCEMENT The Boston University Department of Psychology seeks an Experimental Psychologist with interests in human cognition- attention or memory beginning Fall 1990. The Department seeks candidates with demonstrated excellence in research and teaching. The appointment will be tenure track at the assistant professor level. Candidates should submit a vita, representative reprints and a statement outlining research interests and teaching experience. In addition, they should have three letters of recommendation forwarded. Applications should arrive by February 19, but later applications may be considered. Please write to: Cognitive Search Committee, Department of Psychology Boston University 64 Cummington Street Boston, MA. 02215. U. S. A. Boston University is an Affirmative Action/Equal Opportunity Employer. Women and minority candidates are encouraged to apply. ------------------------------ End of Neuron Digest [Volume 6 Issue 9] ***************************************
hokc_ltd@uhura.cc.rochester.edu (Hok Kiu Chan) (02/08/90)
It is interesting that many of the neural nets we study today are digital, unlike the biological brain, which takes continuous analogy signals. Does anyone know the references for frequency modulated or analog neural nets? I would be very interested to educate myself in this area. Thanks. Victor Hok-kiu Chan