From kaiser@ira.uka.de Wed Oct 20 01:32:59 1993 Received: from iraun1.ira.uka.de by toadflax.cs.ucdavis.edu (4.1/UCD.CS.2.4) id AA16292; Wed, 20 Oct 93 01:32:31 PDT Received: from ira.uka.de by iraun1.ira.uka.de with SMTP (PP) id <25750-0@iraun1.ira.uka.de>; Wed, 20 Oct 1993 09:32:08 +0100 Date: Wed, 20 Oct 93 9:29:19 MET From: Michael Kaiser To: announce@cs.ucdavis.edu Subject: SynWorks announcement Organization: University of Karlsruhe, Germany Department: Computer Science Institute: Real-Time Computer Systems and Robotics Address: P.O.Box 6980, D-76128 Karlsruhe Phone: ++49 721 6084051 Fax: ++49 721 606740 Message-Id: <"iraun1.ira.758:20.09.93.08.32.18"@ira.uka.de> Status: RO Hi Dan, I tried to incorporate your suggestions as good as possible. Don't know about the short NN description - writing an introductory article would require some more space. Therefore I included some references. Hope that's ok. Regards, Michael TITLE SynWorks - A design, training, test, and visualization environment for neural networks. VERSION 2.1, September 1993 AUTHOR Michael Kaiser Rintheimer Strasse 59 D 76131 Karlsruhe, FRG Phone : (+49) (721) 61 18 19 EMail : kaiser@puma.adsp.sub.org kaiser@ira.uka.de DESCRIPTION SynWorks is a fully integrated environment for design, training and test of neural networks. Currently, it supports 11 different network types, the networks' sizes are only limited by the amount of memory available. SynWorks provides instruments to watch and track (to disk) all important network parameters, all kinds of errors and of course, all input and output values provided to or by the network. All network parameters can be edited, and, starting from standard networks, custom networks can be created by adding links and neurons, modifying all functions (learn, transfer, error, evaluation). Input/output behaviour can easily be specified in terms of ranges, sources, and targets. Network display modes include structural, error-related, change-related, weight-related and memory related displays. All displays provide point- and click interfaces to the displayed parameters and are available simultaneously (watch chipmem !). All networks can be printed in four different modes, giving a textual, structural and two different descriptions on the network's weight map. SynWorks also provides a context-sensitive on-line help system that includes descriptions of all actions and important components, linked together via cross-references and easily accessible via buttons and hot-keys. Also, SynWorks features a full AREXX-interface that allows to define macros for internal use as well as using a network as part of a larger system, execute demonstrations etc.. The full version of SynWorks includes five disks with both a 68000 and a 68020/881+ version of SynWorks, several examples, linkable libraries for easy use of networks in other applications (supporting the Amiga and Sun workstations), a tool for data visualization and a set of printed manuals, including a short introduction to neural networks. INTENDED AUDIENCE With SynWorks, people generally interested in neural networks are provided with an easy-to-handle, integrated environment that allows them to experiment with neural networks on their own data. The visualization possibilities of SynWorks make access to the network easy and intuitive. The network's behaviour can be watched and tracked, which in combination with the included examples results in a good insight of how the actual neural network is working. On the other hand, people working on neural network projects and application programmers will find SynWorks powerful enough for most of their needs. Especially in the signal processing field, neural networks are quite attractive. With SynWorks, programmers can create a signal processing or pattern recognition unit on the base of a neural network, and use this network in their application by simple function calls. WHAT ARE NEURAL NETWORKS ? A neural network is a processing device, either an algorithm, or actual hardware, whose design was motivated by the design and functioning of the building blocks of the human brain, such as neurons and connections (axons, dendrits). It is realized as a network of many simple processing units that are regularly interconnected. The connections carry a numerical information, they are "weighted". What makes neural networks very attractive is their ability to "learn" from examples. Most neural networks have some sort of "learning law" which describes how the weights of connections are to be adjusted on the basis of presented patterns. Probably the most popular neural networks are the feedforward networks, with the backpropagation technique/generalized delta rule being the learning law. WHAT CAN NEURAL NETWORKS DO ? In principle, NNs can compute any computable function, i.e. they can do everything a normal digital computer can do. Especially can anything that can be represented as a mapping between vector spaces be approximated to arbitrary precision by feedforward NNs (which is the most often used type). In practice, NNs are especially useful for mapping problems which are tolerant of a high error rate, have lots of example data available, but to which hard and fast rules can not easily be applied. BOOKS ON NEURAL NETWORKS (No completeness intended) Hecht-Nielsen, R. (1990). Neurocomputing. Addison Wesley. Aleksander, I. and Morton, H. (1990). An Introduction to Neural Computing. Chapman and Hall Beale, R. and Jackson, T. (1990). Neural Computing, an Introduction. Adam Hilger, IOP Publishing Ltd. Rumelhart, D. E. and McClelland, J. L. (1986). Parallel Distributed Processing: Explorations in the Microstructure of Cognition (volumes 1 & 2). MIT Press. and lots more ... FEATURES OF SYNWORKS 12 different standard network models 5 different possibilities to display network (simultaneously) 4 different possibilities to print network (standard WB printer, colour printing supported) 24 different instruments to watch and track the network's behaviour 53 AREXX commands give full external control over SynWorks. On-line context sensitive help system, supports keyword tracking. Supports all resolutions higher than or equal to 640 x 400 with at least two bitplanes (incl. A 2024, Productivity, AGA). GUI according to User Interface Style Guide. Version for 68020/68881 and up available. C programming interface available. REQUIREMENTS Amiga 500, 1000, 1200, 2000, 2500, 3000, 4000, 1.5 MByte RAM, Kickstart 2.04 and up. Harddisk recommended, FPU (special program version) highly recommended, Display enhancer/FF/AGA recommended. HOST NAME Aminet: ftp.wustl.edu (128.252.135.4) and its mirrors DIRECTORY /pub/aminet/misc/sci FILE SynWorksDemoAGA.lha PRICE Shareware fee: DM 130/US $80/Students DM 200/US $120/Others DISTRIBUTABILITY Demo version is freely distributable. Full version is shareware, the ftp archive contains a demonstration version with certain features disabled. Users of versions 1.0, 1.1 and 1.2 will be directly informed about upgrade possibilities, all new users can get the full version as described above directly from the author. This version is now fully AGA compatible and contains some new demos. THANKS The maintainers of Aminet.