neuron-request@HPLABS.HP.COM ("Neuron-Digest Moderator Peter Marvit") (05/08/89)
Neuron Digest Monday, 8 May 1989 Volume 5 : Issue 21 Today's Topics: What is the size of LARGE nets ?? Re: Removing noise from EKG signals guassian elimination on sparce matricies (used as an associative mem) Re: guassian elimination on sparce matricies (used as an associative mem) Sorting Re: Sorting E. coli & gradient search Neural Nets in Medical Imaging Re: Neuron Digest V5 #17 Parallel Simulated Annealing / References and Are You Doing It? Share hotel room at IJCNN? Volunteers Wanted for IJCNN Rochester Connectionist Simulator: answers to common questions rochester connectionist simulator available on uunet.uu.net Send submissions, questions, address maintenance and requests for old issues to "neuron-request@hplabs.hp.com" or "{any backbone,uunet}!hplabs!neuron-request" ARPANET users can get old issues via ftp from hplpm.hpl.hp.com (15.255.16.205). ------------------------------------------------------------ Subject: What is the size of LARGE nets ?? From: Rene Bach <mcvax!cernvax!ascom!rene@uunet.uu.net> Organization: Ascom Tech AG, Solothurn, CH Date: 31 Mar 89 07:10:29 +0000 Hi, I am fairly new to this domain. I am curious about the common sizes of useful nets and what size is considered to be LARGE. I would like to know a few things for the LARGE, implemented nets: a) size of input and output layer b) number of nodes (broken down as nodes/hidden layers) and number of connections/node. c) size of training set d) CPU and elapsed training time (and on which hardware) e) a rough measure of performance for that training time and size of training set. If there is interest, I'll summarize to the net. And do not hesitate to let me know whether this is a silly/not relevant question. Cheers Rene Bach Ascom Tech Switzerland rene@ascom.uucp ------------------------------ Subject: Re: Removing noise from EKG signals From: demers@beowulf.ucsd.edu (David E Demers) Organization: EE/CS Dept. U.C. San Diego Date: Fri, 31 Mar 89 17:23:44 +0000 Doug Palmer and Duane DeSieno have a paper, "Removing Random Noise from EKG Signals using a Backpropogation Network". I don't know where it has been published, I have a preprint. Doug Palmer is at HNC, Inc. and Duane DeSieno is at Logical Designs Consulting, Inc. both in the San Diego area. They trained a net to filter out noise and achieved better performance than a finite impulse response filter. Dave DeMers demers@cs.ucsd.edu ------------------------------ Subject: guassian elimination on sparce matricies (used as an associative mem) From: Mike Rynolds <myke@GATECH.EDU> Organization: School of Information and Computer Science, Georgia Tech, Atlanta Date: 04 Apr 89 16:38:29 +0000 If A represents a series of input state vectors and B is a corresponding list of output state vectors, then in the equation AX = B, X is a neural net which can be trained simply by setting it equal to A-1 * B. Since A and B consists of 1's and 0's, and mostly 0's, large matricies can be made managable if they are sparce. I have only been able to find gaussian elimination alg.'s on sparce systems of linear equations of the form Ax = b, where x and b are vectors. Can anyone direct me to where I can find a gaussian elimination alg on sparce systems of the form AX = B? Mike Rynolds School of Information & Computer Science, Georgia Tech, Atlanta GA 30332 uucp: ...!{decvax,hplabs,ncar,purdue,rutgers}!gatech!myke Internet: myke@gatech.edu ------------------------------ Subject: Re: guassian elimination on sparce matricies (used as an associative mem) From: hwang@taipei.Princeton.EDU (Jenq-Neng Hwang) Organization: Princeton University, Princeton NJ Date: Thu, 06 Apr 89 13:50:08 +0000 Instead of using Gaussian elimination type of algorithms for solving the sparse matrices, there have been row-action methods proposed, which are iterative procedures suitable for solving linear systems without any structural assumption on sparseness. One famous example is the Kaczmarz projection method, which can be used to interpret the dynamic behavior of later stage of back-propagation learning, has been widely used in image reconstruction applications. A good tutorial paper is: Yair Censor, " Row-Action Methods for Huge and Sparse Systems and Their Applications," SIAM Review, Vol. 23, No. 4, pp 444-466, October 1981. J. N. Hwang ------------------------------ Subject: Sorting From: Ytshak Artzi - CPS <mailrus!shadooby!accuvax.nwu.edu!tank!eecae!cps3xx!cpsvax!artzi@TUT.CIS.OHIO-STATE.EDU> Organization: Michigan State University, Computer Science Department Date: 11 Apr 89 03:14:24 +0000 Please comment on the following NAIVE algorithm: Let n1, n2,....nk be neurons in a network For every 2 connected neurons (Ni,Nj) we define Ni to "LEFT" and Nj to be "RIGHT" The neurons are initially assigned numeric values (numbers we wish to sort) We activate the network; We let the neurons to exchange information among them "freely" until EVERY "LEFT" contains a smaller value than its "RIGHT" We defined this stage as SORTED NETWORK Questions: 1. Is the computing time predictable ? 2. How can we evaluate the performance of the algorithm (criteria) ? 3. Has anyone done it before ? 4. Does anyone know of a sorting algorithm ? NOTE: if it makes it easier for you, you may assume any Network model, with or without feedback, etc. Thanks. Itzik Artzi, CPS, Michigan State University artzi@cpsvax.cps.msu.edu ------------------------------ Subject: Re: Sorting From: andrew <nsc!andrew@DECWRL.DEC.COM> Organization: National Semiconductor, Santa Clara Date: 11 Apr 89 08:03:56 +0000 > Please comment on the following NAIVE algorithm: > ... For a fully-connected network (as opposed to a 1D string): 1. The network is sorted immediately the numbers are assigned, due to the flexibility in the definition of left, right. Ans: ZERO time. 2. We don't need to bother. 3. Yes, because it's a null operation, and this is occasionally done. 4. Yes. I read Knuth some time ago. Andrew Palfreyman USENET: ...{this biomass}!nsc!logic!andrew National Semiconductor M/S D3969, 2900 Semiconductor Dr., PO Box 58090, Santa Clara, CA 95052-8090 ; 408-721-4788 there's many a slip 'twixt cup and lip ------------------------------ Subject: E. coli & gradient search From: Alex Kirlik <prism!hms2!kirlik@GATECH.EDU> Organization: Center for Human-Machine Systems Research - Ga Tech Date: 11 Apr 89 04:31:41 +0000 The following is taken from Stephen Jay Gould's _Hen's Teeth and Horse's Toes_, W.W. Norton & Co., New York, 1984 (p. 161): "After Berg had modified his microscope to track individual bacteria, he noted that an E. coli moves in two ways. It may "run," swimming steadily for a time in a straight or slightly curved path. Then it stops abruptly and jiggles about--a "twiddle" in Berg's terminology. After twiddling, it runs off again in another direction. Twiddles last a tenth of a second and occur on an average of once a second. The timing of twiddles and the directions of new runs seem to be random unless a chemical attractant exists at high concentration at one part of the medium. A bacterium will then move up-gradient toward the attractant by decreasing the probability of twiddling when a random run carries it in the right direction. When a random run moves in the wrong direction, twiddling frequency remains at its normal, higher level. The bacteria therefore drift toward an attractant by increasing the lengths of runs in favored directions." Not exactly simulated annealing, but close enough for patent infringement? :-) Alex Kirlik UUCP: kirlik@chmsr.UUCP {backbones}!gatech!chmsr!kirlik INTERNET: kirlik@chmsr.gatech.edu ------------------------------ Subject: Neural Nets in Medical Imaging From: bruce raoul parnas <pasteur!sim!brp@UCBVAX.BERKELEY.EDU> Organization: University of California, Berkeley Date: 11 Apr 89 22:00:11 +0000 Can anyone send me any information about/references to the use of neural networks in image processing, specifically in the case of medical imaging. I'm interested in what sorts of things can be done by NNs in recognition of patterns in medical images where there is some a priori knowledge of the scene to be surveyed. if people would prefer to send e-mail, i can post results to the net. thanx, bruce brp@sim.berkeley.edu ------------------------------ Subject: Re: Neuron Digest V5 #17 From: ian@shire.cs.psu.edu (Ian Parberry) Organization: Penn State University Date: Wed, 12 Apr 89 14:02:57 +0000 Bruce, thanks for your interesting reply. I have been away from the net for a while (system installation), sorry if I am a bit out-of-date. >Subject: Re: Re: bottom-up (was Re: NN Question) >From: brp@sim.uucp (bruce raoul parnas) >Date: Wed, 15 Mar 89 17:28:45 +0000 I think we are basically agreed that a statement like "the world is discrete" or "the world is analog" gives us little reason to model neural networks as discrete or analog. >>Real numbers, continuous functions etc., are abstractions which help >>us deal with the fact that the number of discrete units is larger >>than we can deal with comfortably. > >right. and in most physical systems we may, for our understanding, treat them >as essentially analog since we simply can't deal with the complexity presented >by the true (?) discrete nature. I'm not convinced. Computational complexity theory gives us tools for dealing with discrete resources (time, memory, hardware) which are too large to handle individually. There is no need to treat them as continuous. >>There are (at least) two objections to the classical automata- >>theoretic view of neural systems. One is that neural systems >>are not clocked (I presume that this is what you mean by >>"continuous time"), and that neurons have analog behaviour. > >that is precisely what i meant. neurons each evolve on their own, independent >of system clocks. Yes? I didn't think the evidence was in on that. I recently heard of a paper that claimed a large amount of synchronicity in neuron firings. I don't remember the author. I'll send you email if I remember. >i believe that a system clock would be more of a hindrance that a help. >studies with central pattern generators and pacemaker activity (re: the heart) >show clearly that system clocks are not unavailable. if evolution had found >a neural system clock advantageous, one could have been created. i feel, >however, that the continuous-time evolution of neural systems imbues them >with their remarkable properties. You are entitled to your opinion. You are reasoning by analogy here. Could there REALLY be a wetware system clock? You may be missing implementation details that make it impossible. For example, could the correct period (milliseconds) be achieved? And could it be communicated reliably and in small hardware to all neurons? I think the remarkable properties of neural networks come from other sources; or perhaps we have different definitions of "remarkable". Here is another way of looking at it. When one neuron fires and its neighbour is not receptive (building up charge) there is a fault. Faults are relatively infrequent (receptive time is larger than nonreceptive time). The architecture is fault-tolerant. That's why we observe that the brain is fault-tolerant when some of its neurons are destroyed. It has to be in order to get around the lack of system clock. Neural architectures are better at fault-tolerance than von-Neumann ones (at least, we can prove this when the thresholding is physically separated from the summation of weights, as seems to be the case for biological neurons). >I don't think that such a fine level of precision is necessary in neural >function, i.e. six places would likely be enough. but since digital circuitry >is made actaully from analog circuit elements limited to certain regions of >operation, why go to this trouble in real neural systems when analog seems >to work just fine? If six decimal places is enough, then we can model everything as integers. Why do this? It is easier to analyze. Combinatorics is easier than analysis (despite Hecht-Nielson's claim in the first San Diego NN conference that the opposite is true). I don't care if the real neural systems seem to behave in an analog fashion. If it seems that the _computationally important_ things going on are really discrete (and you seem to have agreed that this is the case), then our model should reflect this. I'm not necessarily saying that we should _build_ them that way. That's another question. But perhaps we ought to _think_ of them that way. To use an analogy, we don't usually think of a computer as having infinite memory, but it certainly helps to program them as if it were the case. For a complexity theorist, infinite means "adequate for day-to-day use". This is where the classical attack on theoretical computer science (my TRaSh-80 is not a Turing machine) breaks down. I think that, despite the bad press that theoretical computer science gets from some NN researchers (I've heard many unprofessional statements made in conference presentations by people who should know better), complexity theory has something to contribute. So do other disciplines. I'm just a little tired of people closing doors in my face. It has become fashionable to disparage TCS (following the bad examples mentioned three sentences ago). Sorry if my knee-jerk reaction to your posting was a little harsh. Ian Parberry "The bureaucracy is expanding to meet the needs of an expanding bureaucracy" ian@theory.cs.psu.edu ian@psuvax1.BITNET ian@psuvax1.UUCP (814) 863-3600 Dept of Comp Sci, 333 Whitmore Lab, Penn State Univ, University Park, Pa 16802 ------------------------------ Subject: Parallel Simulated Annealing / References and Are You Doing It? From: "Dan R. Greening" <squid!dgreen@LANAI.CS.UCLA.EDU> Organization: UCLA Computer Science Department Date: 13 Apr 89 23:14:42 +0000 Simulated annealing is often used for combinatorial optimization and other hard problems. It uses thermodynamic properties as a metaphor for solving these problems. Simulated annealing is used for nearly every VLSI CAD problem under the sun: placement, routing, logic optimization, circuit delay. Plus some vision problems, neural network weight-setting, a huge collection of NP complete problems, etc. That's why the crossposting list is so big. I probably missed a few, too :-). I am looking for references AND people who have implemented simulated annealing applications on parallel processors. There are some difficult and interesting problems in extending simulated annealing to parallel processors. If you send parallel simulated annealing references to me via e-mail, I will compile a bibliography and share it with all contributors. I'm also interested in hearing from people who are actively WORKING in parallel simulated annealing, and will set up a mailing list if there is enough interest. Any leads you can give me to parallel simulated annealing researchers who may not read news (but have e-mail connections), would also be greatly appreciated. Dan Greening dgreen@cs.ucla.edu 213-472-4898 308 Westwood Plaza, #117 Los Angeles, CA 90024-1647 ------------------------------ Subject: Share hotel room at IJCNN? From: mv10801@uc.msc.umn.edu Date: Wed, 03 May 89 14:25:17 -0500 I would like to find a roommate to share at hotel room with at the IJCNN conference in Washington DC in June. Male non-smoker preferred. If you're interested, please contact me by e-mail or by phone. Thanks! -Jonathan Marshall mv10801@uc.msc.umn.edu Center for Research in Learning, Perception, and Cognition 205 Elliott Hall University of Minnesota 612-331-6919 (eve/weekend/msg) Minneapolis, MN 55455 612-626-1565 (office) ------------------------------ Subject: Volunteers Wanted for IJCNN From: neilson%cs@ucsd.edu (Robert Hecht-Nielsen) Date: Wed, 03 May 89 15:43:55 -0700 Request for volunteers for the upcoming International Conference on Neural Networks (IJCNN) June 18 - June 22 Requirements: In order to receive full admission to conference and the proceedings, you are required to work June 19 - June 22, one shift each day. On June 18 there will be tutorials presented all day. In order to see a tutorial, you must work that tutorial. See the information below on what tutorials are being presented. Shifts: There are 3 shifts: Morning, afternoon and evening. It is best that you work the same shift each day. Volunteers are organized into groups and you will, more than likely, be working with the same group each day. This allows at great deal of flexibility for everyone. If there is a paper being presented at the time of your shift, you can normally work it out with your group to see it. Last year I had no complaints from any of the volunteers regarding missing a paper which they wanted to view. Tutorials: The following tutorials are being presented: 1) Pattern Recognition - Prof. David Casasent 2) Adaptive Pattern Recognition - Prof. Leon Cooper 2) Vision - Prof. John Daugman 4) Neurobiology Review - Dr. Walter Freeman 5) Adaptive Sensory Motor Control - Prof. Stephen Grossberg 6) Dynamical Systems Review - Prof. Morris Hirsch 7) Neural Nets - Algorithms & Microhardware - Prof. John Hopfield 8) VLSI Technology and Neural Network Chips - Dr. Larry Jackel 9) Self-Organizing Feature Maps - Tuevo Kohonen 10) Associative Memory - Prof. Bart Kosko 11) Optical Neurocomputers - Prof. Demitre Psaltis 12) Starting a High-Tech Company - Peter Wallace 13) LMS techniques in Neural Networks - Prof. Bernard Widrow 14) Reenforcement Learning - Prof. Ronald Williams If you want to work the tutorials, please return to me you preferences from 1 to 14 (1 being the one you want to see the most). Housing: Guest housing is available at the University of Maryland. It is about 30 minutes away from the hotel, but Washington D.C has a great "metro" system to get you to and from the conference. The cost of housing per night is $16.50 per person for a double room, or $22.50 for a single room. I will be getting more information on this, but you need to sign up as soon as possible as these prices are quite reasonable for the area and the rooms will go quickly. General Meeting: A general meeting is scheduled at the hotel on Saturday, June 17, around 6:00 pm. You must attend this meeting! If there is a problem with you not being able to make the meeting, I need to know about it. When you contact me to commit yourself officially, I will need from you the following: 1) shift preference 2) tutorial preferences 3) housing preference (University Housing?) To expedite things, I can be contacted at work at (619) 573-7391 during 7:00am-2:00pm west coast time. You may also leave a message on my home phone (619) 942-2843. Thank-you, Karen G. Haines IJCNN Tutorials Chairman neilson%cs@ucsd.edu ------------------------------ Subject: Rochester Connectionist Simulator: answers to common questions From: rochester!bukys@RUTGERS.EDU Organization: U of Rochester, CS Dept, Rochester, NY Date: 07 Apr 89 17:41:57 +0000 From: bukys =============================================================================== ANSWERS TO COMMON QUESTIONS REGARDING THE ROCHESTER CONNECTIONIST SIMULATOR Fri Apr 7 12:34:10 EDT 1989 Liudvikas Bukys <bukys@cs.rochester.edu> =============================================================================== WHAT IS THE SIMULATOR? The Rochester Connectionist Simulator is a tool that allows you to build and experiment with connectionist networks. It provides the basic mechanism for running a simulation (iterate through all units, call functions, update values). It also provides a graphic interface that lets you examine the state of a network. It also provides convenient facilities for defining and manipulating your network: names for units, set manipulation, etc. It also has a dynamic loading facility, so you can compile and load new functions on the fly, and to allow to customize the simulator by adding your own commands. There is also a library to help you implement back-propagation networks. The Simulator does come with a few simple canned examples, but does not provide a lot of the latest greatest gizmos that researchers have dreamed up. You should think of the simulator as a network programmer's tool -- you have the tool, but you have to know what to do with it. WHAT MACHINES DOES IT RUN ON? We have run it here on VAX and SUN-3 (SunOS 3.x) systems here. It has been made to run on SUN-4 systems without too much trouble. It's a little boring without the graphic interface, though, and... The graphic interface only runs on Sun workstations under the SunView window system. WHAT OPERATING SYSTEM DEPENDENCIES ARE THERE? The code is generally pretty generic C, so running it on strange machines shouldn't be too much trouble, except for the dynamic loading modules. The Simulator should feel at home on most Berkeley Unix (BSD 4.2/4.3) based operating systems. In particular, if your system has an a.out.h file in /usr/include, it'll probably install easily. If your machine is based on AT&T System V, your object files are probably COFF format, and some changes need to be made. We are not running SunOS 4.0 here yet, so it is only known to run on SunOS 3.x systems. Again, the dynamic loading routines need a few changes to run under this operating system. If you have the simulator running on either System V or SunOS 4.0, please send me your patches, so that I can share them with everyone else. DO YOU HAVE AN X VERSION OF THE GRAPHIC INTERFACE? Someone here worked on a version of the graphic interface using X11 and the HP toolkit. It was close to working, but not quite debugged (as far as I can tell). Someone else (who will remain nameless for the moment) ported that code to X11 and the Athena toolkit. This code will probably be available soon. Stay tuned to this mailing list. IS THERE A VERSION AVAILABLE FOR MULTI-PROCESSORS? Well, yes and no. Version 4.0 was ported to the BBN Butterfly I. Unfortunately, this was one of the first things we did with the Butterfly, at a time when the programming environment was quite crude. (For those who know the Butterfly: this was before the existence of the Uniform System or the Streams library.) Getting it to run now is still possible, but it would require digging up a pile of little utility daemons and libraries and so forth, and, chances are, unless you are really serious about it, you'll give up! Perhaps only the original author (Mark Fanty) or I (Liudi Bukys) would really be able to do it, and we're both doing better things right now. If, after reading all this, you still want to try, contact me, and I'll gather the package up and send it to you. P.S. The back-propagation library doesn't run under this old Butterfly version. Of course, the current environment (Mach) on modern Butterflies is much better, so it's possible that someone here will port it there some time. To my knowledge, no one has ported a parallelized version to any of the other common multiprocessors (Sequent, Alliant, Encore, Inmos/Transputer, Intel/HyperCube). =============================================================================== Liudvikas Bukys <simulator-request@cs.rochester.edu> ------------------------------ Subject: rochester connectionist simulator available on uunet.uu.net From: Liudvikas Bukys <rochester!bukys@CU-ARPA.CS.CORNELL.EDU> Organization: U of Rochester, CS Dept, Rochester, NY Date: 21 Apr 89 14:39:56 +0000 A number of people have asked me whether the Rochester Connectionist Simulator is available by uucp. I am happy to announce that uunet.uu.net has agreed to be a redistribution point of the simulator for their uucp subscribers. It is in the directory ~uucp/pub/rcs on uunet: -rw-r--r-- 1 8 11 2829 Jan 19 10:07 README -rw-r--r-- 1 8 11 527247 Jan 19 09:57 rcs_v4.1.doc.tar.Z -rw-r--r-- 1 8 11 9586 Jul 8 1988 rcs_v4.1.note.01 -rw-r--r-- 1 8 11 589 Jul 7 1988 rcs_v4.1.patch.01 -rw-r--r-- 1 8 11 1455 Apr 19 19:18 rcs_v4.1.patch.02 -rw-r--r-- 1 8 11 545 Aug 8 1988 rcs_v4.1.patch.03 -rw-r--r-- 1 8 11 837215 May 19 1988 rcs_v4.1.tar.Z These files are copies of what is available by FTP from the directory pub/rcs on cs.rochester.edu (192.5.53.209). We will still send you, via U.S. Mail, a tape and manual for $150 or just a manual for $10. If you are interested in obtaining the simulator via uucp, but you aren't a uunet subscriber, I can't help you, because I don't know how to sign up. Maybe a note to postmaster@uunet.uu.net would get you started. Liudvikas Bukys <simulator-request@cs.rochester.edu> ------------------------------ End of Neurons Digest *********************