neuron-request@HPLMS2.HPL.HP.COM ("Neuron-Digest Moderator Peter Marvit") (06/01/90)
Neuron Digest Thursday, 31 May 1990 Volume 6 : Issue 37 Today's Topics: Special Issue on Neural Networks Neuron Digest Submission UPDATED program info for: 6/16 NEURAL NETS FOR DEFENSE Conference TIME-SENSITIVE - DoD small Business Innovation Research Program 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: Special Issue on Neural Networks From: Alexander Linden <hplabs!gmdzi!al> Date: Thu, 31 May 90 09:31:58 -0200 Special Issue on Neural Networks in Parallel Computing (To appear in August) This special issue focuses on the third generation of neural networks, which can be characterized as being heterogeneous, modular and asynchronous. Contents: H. Muehlenbein: Limitations of Multilayer Perceptrons: Towards Genetic Neural Networks F. Smieja: The Geometry of Multilayer Perceptron Solutions H. Muehlenbein J. Kindermann: Inversion of Neural Networks by Gradient Descent A. Linden T. E. Lange: Simulation of Heterogeneous Neural Networks on Serial and Parallel Machines A. Singer: Implementations of Artificial Neural Networks on the Connection Machine X. Zhang: The Backpropagation Algorithm on Grid and Hypercube et al. Architectures M. Witbrock: An Implementation of Backpropagation Learning on GF11, M. Zagha a large SIMD Parallel Computers D. Whitley: Genetic Algorithms and Neural Networks: Optimizing et al. Connections and Connectivity M. Tenorio: Topology Synthesis Networks: Self Organization of Structure and Weight Adjustment as a Learning Paradigm K. Obermayer: Large Scale Simulations of Self-Organizing Neural Networks on Parallel Computers: Application for Biological Modelling R. Kentridge: Neural Networks for Learning in the Real World: Representation, Reinforcement and Dynamics - ------------------------------------------------------------------------- HOW TO ORDER: The publisher is offering a special service. Copies of this issue at a price of $25 can be obtained from Dr. F. van Drunen Elsevier Science Publishers Mathematics and Computer Science Section P.O. BOX 103 1000 AC Amsterdam The Netherlands FAX: +31-10-5862-616 - ------------------------------------------------------------------------- Copies of the first three papers can be got from GMD c/o Sekretariat Z1.HLRZ P.O. BOX 1240 D-5205 Sankt Augustin 1 West Germany FAX +49 - 2241 - 142618 Heinz Muehlenbein ------------------------------ Subject: Neuron Digest Submission From: erik%adams.llnl.gov@lll-lcc.llnl.gov (ERIK JOHANSSON) Date: Thu, 31 May 90 09:11:42 -0700 I have subscribed to the digest for a couple of months now, but have not submitted anything. Judging by the recent traffic, I feel now is the time. I am a research engineer at Lawrence Livermore National Laboratory, working in the areas of signal and image processing and neural network applications. We have a neural network working group of 10 - 12 people which meets on a regular basis. Our applications and areas of interest include: pattern recognition, parallel implementations of NNs using systolic arrays, signal detection and estimation, image data analysis, chemical data analysis (chemometrics), optimization methods, NN theory, and architectures. I am working on optimization methods and chemical data analysis. In December of last year, I successfully applied the conjugate gradient optimization method to backpropagation networks. The results were excellent, and I'm almost finished with a paper comparing back propagation with several different forms of the conjugate gradient method (the paper would have been completed some time ago, but my work schedule would not permit it). I should have it finished in about 1 - 2 weeks, and as soon as I get a publication release from the lab, I will make preprints available on the net. I am briefly summarizing the results in this submission. To begin with, I needed some kind of metric to use in comparing the performance of conventional backpropagation (with momentum) against the conjugate gradient method. I agree with recent statements in the digest that the number of cycles or iterations is not a valid metric, since in backpropagation the training set is passed through the network once per iteration, whereas in the conjugate gradient method the training set can be passed through several times in a single iteration. The majority of the CPU time required to train a backpropagation network is spent computing the error function and the gradient of the error function. Computing the error function requires forward propagating the entire training set through the network. Computing the gradient requires backpropagating the outputs from each pattern in the training set (computed during the forward propagation) through the network. Conventional backpropagation requires one forward propagation and one backpropagation per iteration. The conjugate gradient method will have several forward and back propagations at each iteration. The number of forward and back propagations may or may not be the same, depending on the type of line search (1-D minimization) used. Some line searches use both the function value and the gradient at each step, some use only the function value, and some use a combinations of the two. It seems to me that since most of the CPU time is spent computing the error function and gradient, the logical metric for comparison is the total number of function calculations AND the total number of gradient calculations required to train a network. Using this metric, I have compared backpropagation with the conjugate gradient method. For the conjugate gradient method, I used a cubic approximation line search which computes function values and gradient values at each step. Therefore, the number of function and gradient calculations are identical. Since this is also the case with conventional backpropagation, I have reduced the original metric to just 'function evaluations' for simplicity. I chose the parity problem as a benchmark test. I realize many people feel the parity problem is somewhat trivial; however, the error surfaces are very complex (I've looked at slices along search directions, and the error functions are EXTREMELY nonlinear). This nonlinearity presents quite a challenge to optimization techniques, and hence makes for a nice benchmark. I tested both one and two hidden layer networks on 3, 4, and 5 bit parity problems (i.e. 3-3-1, 3-3-3-1, 4-4-1, 4-4-4-1, 5-5-1, and 5-5-5-1 architectures). I used a full training set for each test (2^nth patterns), with a stopping criterion of 1.e-6 for the normalized system error (the average of the pattern errors). In addition, the number of function evaluations were limited to 50000 - the optimization terminated if this limit was reached. The weights were initialized to random numbers uniformly distributed between -0.5 and +0.5. The backpropagation tests were performed using several combinations of step size and momentum. I tested several versions of the conjugate gradient method: Fletcher-Reeves, Polak-Ribiere, Hestenes-Steifel, and Shanno's method. Each method was tested using several values of the linesearch termination critrerion (too complex to explain here - the paper will explain in detail). In addition, I ran the tests using pure steepest descent with a linesearch. The results are summarized below; only the best results for each test are shown. We found the conjugate gradient method to be at least 15 times faster than conventional backpropagation, and in most cases many times faster. In addition, our experience involving problems of considerable size has shown that the conjugate gradient method is much faster in general than conventional backpropagation. Test Results: BP - Backpropagation FR - Fletcher - Reeves PR - Polak - Ribiere HS - Hestenes - Steifel SH - Shanno SD - Steepest descent 3 bit parity, 1 hidden layer (3-3-1): Method # Func. Evals Error speedup over BP FR 264 9.48e-7 7.0 PR 123 8.29e-7 15.0 HS 121 7.02e-8 15.2 SH 191 5.45e-7 9.6 SD 616 9.91e-7 3.0 BP 1843 9.98e-7 (BP stepsize 0.9, momentum 0.9) 3 bit parity, 2 hidden layers (3-3-3-1) Method # Func. Evals Error speedup over BP FR 372 7.47e-7 8.3 PR 200 7.87e-7 15.4 HS 239 9.11e-7 12.9 SH 275 7.33e-7 11.2 SD 1148 9.69e-7 2.7 BP 3078 9.81e-7 (BP stepsize 0.9, momentum 0.9) 4 bit parity, 1 hidden layer (4-4-1) Method # Func. Evals Error speedup over BP FR 1617 9.55e-7 NA PR 461 9.39e-7 NA HS 306 8.59e-7 NA SH 2079 8.71e-7 NA SD 5505 9.99e-7 NA BP 50000 1.60e-2 (did not converge) 4 bit parity, 2 hidden layers (4-4-4-1) Method # Func. Evals Error speedup over BP FR 737 9.88e-7 18.3 PR 401 7.01e-7 33.6 HS 429 9.09e-7 31.4 SH 560 9.96e-7 24.0 SD 1840 9.98e-7 7.3 BP 13462 9.90e-7 (BP stepsize 0.7, momentum 0.7) 5 bit parity, 1 hidden layer (5-5-1) Method # Func. Evals Error speedup over BP FR 2145 9.74e-7 11.7 PR 1966 7.76e-7 12.8 HS 3249 8.89e-7 7.7 SH 750 8.97e-7 33.4 SD 11894 9.99e-7 2.1 BP 25087 9.98e-7 (BP stepsize 0.9, momentum 0.9) 5 bit parity, 2 hidden layers (5-5-5-1) Method # Func. Evals Error speedup over BP FR 1561 9.42e-7 NA PR 1343 9.91e-7 NA HS 1165 9.56e-7 NA SH 1890 9.86e-7 NA SD 9500 9.99e-7 NA BP 50000 8.00e-2 (did not converge) In all of the above tests, some form of the conjugate gradient method was always at least 15 times faster than backpropagation. In most cases, most notably those where backprop did not converge, it is many times faster. The conjugate gradient method can get stuck in local minima; however, the algorithm can be modified to detect this and use a simple pattern search to get out of the minimum, and then continue with the conjugate gradient method. In addition, my experience using the algorithm on pattern recognition problems has shown that when the algorithm gets stuck, it is usually due to a large "flat" plateau in the error surface where the gradient becomes quite small, not a well defined local minimum. Again, the use of a pattern search (a systematic search through the error space) can resolve this problem. In general, I find the conjugate gradient method to be quite superior to conventional backpropagation. Indeed, from an optimization viewpoint, the idea of using a fixed step size is not a good one: the move taken can either be so small that it would take an exceedingly long time to converge, or so large that the minimum is missed and the algorithm oscillates about the minimum, converging very slowly. The linesearch in the conjugate gradient method corrects this problem by finding the minimum along a search direction at each iteration to an accuracy specified by the user. As with many complex numerical algorithms, the conjugate gradient method may require the use of double precision variables (this is problem dependent), but the speedup is well worth the small increase in computation. The paper, which will be completed shortly, has a detailed tutorial derivation of the conjugate gradient method, an explanationm of its application to the backpropagation learning problem, and a complete listing of all the test results. I look forward to any comments or questions the digest readers may have. Sincerely, Erik M. Johansson Lawrence Livermore National Laboratory PO Box 808, L-156 Livermore, CA 94550 erik@adams.llnl.gov (415) 423-9255 Disclaimer: The opinions expressed herein are my own and do not necessarily represent the views of Lawrence Livermore National Laboratory, the University of California, or the U.S. Government. ------------------------------ Subject: UPDATED program info for: 6/16 NEURAL NETS FOR DEFENSE Conference From: neuron-request@hplabs.hpl.hp.com Date: Thu, 31 May 90 10:43:28 -0700 [[ Editor's Note: Comments directly relevant to the charter of this Digest - artificial and natutal neural networks - are always welcome. A reminder, however, that general pro-/anti- defense debates are appropriate in a different forum. -PM ]] UPDATED Program and Registration information for: The Second Annual Conference on NEURAL NETWORKS FOR DEFENSE Conference Chair: Prof. Bernard Widrow June 16th, 1990: ** The day before IJCNN ** San Diego Marriot, San Diego, CA "Neural Networks for Defense" is organized to encourage and promote the transfer of neural network technology to practical defense- oriented applications. PROGRAM (June 16th, 1990): ========================== MORNING:__________________________________________________________________ 8:15am-8:45am Mark A. Gluck (Stanford University): "Opening remarks" Robert Kolesar (Deputy Director for Adv. Technology, Naval Ocean Systems Center) "Defense funding of neural networks: A programmatic overview of 6.1 -> 6.3 efforts" 8:45-10:30 PANEL SYMPOSIUM: INTERNAL DOD LABORATORIES______________________ Steven Speidel (Naval Ocean Systems Center) "A neural target locator" Steven Anderson (Captain, USAF; Air Force Weapons Laboratory, KAFB,NM) "Neural networks for signal processing and pattern recognition" Steven K. Rogers (Major, USAF; Air Force Institute of Technology, WPAFB, OH) "Artificial Neural Networks for Automatic Target Recognition" David Andes (Director of Neural Network R&D, Naval Weapons Center, China Lake) "Artificial neural computing at the Naval Weapons Center" 10:30-11:00am COFFEE BREAK 11:00-12:30 PANEL SYMPOSIUM: SBIR SUPPORT OF NN R&D:_________________________ Craig Will (Editor, Neural Network Review) "An overview of neural network research in the SBIR program" Vincent D. Schaper (Navy SBIR Manager) "The DoD SBIR program" Luis Lopez (US Army Strategic Defense Command) "U.S. Army Strategic Defense Commands' SBIR Neural Network Programs" Robert L. Dawes (President, Martingale Research Corporation) "Observations on the SBIR program by a successful participant" James Johnson (Regional Vice President, Netrologic) "SBIRs: A contractor and government perspective" 12:30-2:00 AFTER-LUNCH SPEAKER: Bernard Widrow (Stanford University) 2:00-4:00 2:00-4:00 SESSION: PROGRESS IN DEFENSE APPLICATONS:__________________________ Edward Rosenfeld (Intelligence Newsletter) "Overview of Industry efforts in Neural Networks for Defense" David Hamilton (Senior Development Engineer, Raytheon Submarine Signal Div.) "Neural Network Defense Applications within Raytheon" Robert North (President, HNC, Inc.) "Neural Network Defense Applications at HNC" Rich Peer (Senior Manager, McDonnel Douglas) "Neural Network Applications at McDonnell Douglas" Donald F. Specht (Senior Scientist, Lockheed Research Laboratory) "Hardware Implementation of Neural Networks" Joseph Walkush (SAIC) "Neural Networks for Defense and Security at SAIC" 4:00-4:30 -- COFFEE BREAK 4:30-5:30 PANEL SYMPOSIUM: FORGING TRANSITIONS BETWEEN UNIVERSITIES AND FOR ADVANCED APPLICATIONS OF NN FOR DEFENSE_______________________ Thomas McKenna (Scientific Officer, Office of Naval Research" "Navy Transition Paths from Basic to Applied Research" James Anderson (Prof. of Cognitive Science, Brown University) "Highs and lows: A case study" Terrence Sejnowski (Institute for Neural Computation, UC, San Diego/Salk Institute) "Case history of a successful university-industry cooperative venture" REGISTRATION: ============= This meeting is UNCLASSIFIED but limited to those with an explicit "need-to-know" and a clear professional commitment to the defense and security interests of the United States. ***** ATTENDANCE IS STRICTLY LIMITED TO U.S. CITIZENS ONLY. ********* NOTE: Special Registration Fee Discounts for DoD Employees & University Scientists working on DoD 6.1 Research For further information, or to register, contact: ------------------------------------------------- Lynne Mariani, Registration Coordinator Neural Networks for Defense 500 Howard St. San Francisco, CA 94105 Phone: (415) 995-2471 FAX: (415) 995-2494 ------------------------------ Subject: TIME-SENSITIVE - DoD small Business Innovation Research Program From: will@ida.org (Craig Will) Date: Thu, 31 May 90 16:29:36 -0400 Please get this out quickly because if its nature. Tks, Craig Will Department of Defense Small Business Innovation Research Program The U. S. Department of Defense has announced an unusual mid-year solicitation for proposals for the Small Business Innovation Research (SBIR) program. The list of topics included in these soliciations is of general interest in that they reflect the increasing level of interest by mil- itary agencies in neural network technology and their perception of the kinds of applications they are interested in solving with the technology. In the current solicitation only the Army, Navy, and DARPA are participating; the Air Force and SDIO are not. There are 11 topics specifically targeting neural networks, and at least 2 more topics that specifically mention neural networks as possible approaches that might be used. The program is in three Phases. Phase I awards are essentially feasibility studies of 6 months and with a dollar amount of about $50,000, intended for a one-half person-year effort. Phase I contrac- tors compete for Phase II awards of 2 years in length and up to $500,000, intended for 2 to 5 person-years of effort. Phase III is the commercial application phase of the research. Proposals must be no longer than 25 pages in length, including the cover sheet, summary, cost proposal, resumes and any attachments. Deadline for proposals is July 2, 1990. Principal investigators must be employees (50% or more time) of small business firms. The program encourages small businesses to make use of university-based and other consultants when appropriate. The topics listed below are those that are specifically targeted to or mention neural networks specifically as possible approaches: Many other topics describe application and problem areas that could also be solved with these approaches. A90-215 Neural Network Based Classification Demonstration of Vehicle from Laser Radar and Infrared Data. (Exploratory Develop- ment). Development of methods to classify military vehicles (e.g., distinguish a tank from a truck).using laser radar and infrared images. Phase I involves development of a neural network vehicle classifier based on laser rader data. Phase II involves integrating laser radar and infrared imagery together ``to demonstrate a multi- sensor classifier showing high probability of classification and low false-alarm rate." A90-227 Application of Neural Networks to Command and Control. (Exploratory Development). Apply neural networks to ``information processing and decision making in a command and control operating environment." Neural networks are seen as having ``promise for pro- viding significant improvements in reaction times by providing quantum leaps in the ability to quickly process information and perform deci- sion aid tasks". Phase I involves development of a plan and demons- tration software. Phase II involves the ``development and testing of a working system in a field environment." A90-236 VHSIC Application to Neural Network Systems. (Explora- tory Development). Investigate application of Very High Speed Integrated Circuit technology to the problem of electronically imple- menting neural networks that require very large connectivity. Phase I involves ``identification of specific very large Neural Networks which could be implemented using VHSIC technology, and development of a demonstratable design". Phase II involves ``fabrication and test of a hardware brassboard for one or more specific applications". A90-246 Neural Network Sensor Fusion for Apache Escort Jammer Countermeasure Systems. (Exploratory Development). Develop a system using neural networks that can provide improved awareness of the situation and management of electronic countermeasures for a pilot in an electronic warfare environment. The system should provide improved accuracy and/or increased processing speed. Phase II involves defin- ing the network architecture and training, and possibly simulate and test the network in a demonstration setting. Phase II involves simu- lating the network ``in a manner than approaches real-time perfor- mance," test it, embed the net in an Apache Escort system [that manages countermeasures] and demonstrate its capabilities. A90-428 Neural Network Software / Hardware for Directed and Kinetic Energy Anti-satellite (ASAT) Weapons System. (Basic Research). Development of ``new and innovating neural network algo- rithms and architectures that will aid in developing a real-time, economical and reliable kinetic and directed energy antisatellite (ASAT) weapons system." Problems include ``weapons pointing, beam control, acquisition, tracking, sensor focal planes, signal and data processing, guidance and control algorithms, control of cryocoolers, [and] array image processing". Phase I involves demonstration of con- cept by simulation or prototype development. Phase II involves ``incorporating the principle developed in Phase I into a prototype" or showing proof of feasibility for a demonstration phase. N90-317 Neural Network Applications for Nondestructive Inspec- tion of Aircraft. (Research). Develop new approaches to automatic inspection of aircraft that is nondestructive. X-ray, ultra-sonic, eddy-current, and acoustic sensors are typically used to detect such flaws as cracks, voids, porosity, incudisions, delaminations, pits, corrosion, etc. Neural networks are seen as a way of integrating this sensor data so as to recognize specific flaws. Neural network approaches for ``robotic sensor placement" are also of interest. N90-351 Artificial Intelligence and Neural Network Technologies for Mission Planning and Execution Applications. (Exploratory Development). Application of AI and neural network techniques to help automate or assist in the planning and execution of missions. Goal is to develop techniques that can result in a fielded system within ``five to fifteen years". Phase I involves assessment of mission planning and identification of promising technologies. Phase II involves building a demonstration system using the concepts developed in Phase I. N90-372 Neural Network Applications to Flight Control (Explora- tory Development). Investigate potential for using neural networks to help stabilize a high-performance aircraft in flight that is subject to changing environmental conditions as well as instabilities result- ing from its own dynamics. Phase I involves development and demons- tration by simulation of a network architecture that can stabilize an aircraft. Phase II expands to ``provide robust control and stabiliza- tion features in a distributed neural network having excellent sur- vivability and fault tolerant properaties." N90-384 LSI (Large System Integrated) Neural Networks for Asso- ciative Memory Arrays. (Advanced Development). Investigate neural network architectures and hardware implementation techniques for contstruction of ``an associative memory array" based on artificial neural networks to ultimately achieve 10 ** 11 and 10 ** 12 intercon- nects / second" for application to video and audio ``matching" prob- lems. Phase I involves investigating materials, devices, architec- tures, and modeling. Phase II involves ``a technology demonstration illustrating the several orders of magnitude improvement offered by the physical use of VLSI associative memory arrays based on ANNs." DARPA 90-115 Unique Applications for Artificial Neural Networks. (Exploratory Development). Identification and development of applica- tions that can show ``outstanding potential to demonstrate particular advantages of artificial neural networks...in systems that perform challenging tasks that are at or beyond the limits of capability of conventional information processing systems." Applications that help discover ``important unusual and under-recognized `niches'" for neural networks are particularly sought. Phase I involves a providing a con- ceptual design and laboratory demonstration. Phase I involves building a compact prototype system. DARPA 90-124 Artificial Neural Network Target Recognition Demonstration. (Basic Research). Develop hardware for implementing a specific neural network algorithm that has been developed by the Army for object extracting-classifying pixels in an image into candidate regions suggesting objects. Details of the algorithm will be fur- nished by DARPA ``as required". Phase I involves design and demons- tration of a candidate hardware approach that shows scalability and real-time operation. Phase II involves building a full-scale, real- time hardware system that can process real images as a laboratory demonstration. `` Topics A90-430 and A90-473 also specifically mention the possi- blity of using neural network approaches, while many other topics are also presumably candidates. For more details on the July, 1990 soliciation obtain a copy of the SBIR Program Solicitation book (229 pages in length) from the Defense Technical Information Center: Attn: DTIC/SBIR, Building 5, Cameron Station, Alexandria, Virginia 22304-6145. Telephone: Toll- free, (800) 368-5211. For Virginia, Alaska, Hawaii: (202) 274-6902. Craig A. Will Computer and Software Enginering Division Institute for Defense Analyses Alexandria, VA will@ida.org ------------------------------ End of Neuron Digest [Volume 6 Issue 37] ****************************************