********************************************************************** SNNS (Stuttgart Neural Network Simulator) Version 4.2 available ********************************************************************** The new version 4.2 of SNNS is available now. ********************************************************************** New features of SNNSv4.2 ********************************************************************** Version 4.2 of SNNS features the following improvements and extensions over the earlier version 4.1: * greatly improved installation procedure * pattern remapping functions introduced to SNNS * class information in patterns introduced to SNNS * change to all batch algorithms: The learning rate is now divided by the number of patterns in the set. This allows for direct comparisons of learning rates and training of large pattern files with BP-Batch since it doesn't require ridiculous learning rates like 0.0000001 anymore. * Changes to Cascade-Correlation: -- Several modifications can be used to achieve a net with a smaller depth or smaller Fan-In. -- New activation functions ACT_GAUSS and ACT_SIN -- The backpropagation algorithm of Cascade-Correlation is now available in an off-line and a batch version. -- The activations of the units can be cached. The result is a faster learning for nets with many units. On the other hand, the necessary memory space will increase for large pattern sets. -- Changes in the 2D-display, the hidden units are displayed in layers, the candidate units are placed on the top of the net. -- validation now possible -- automatic deletion of candidate units at the end of training. * new meta learning algorithm TACOMA. * new learning algorithm BackpropChunk. It allows chunkwise updating of the weights as well as selective training of units on the basis of pattern class names. * new learning algorithm RPROP with weight decay. * algorithm "Recurrent Cascade Correlation" deleted from repository. * the options of adding noise to the weights with the JogWeights function improved im multiple ways. * improved plotting in the graph panel as well as printing option. * when standard colormap is full, SNNS will now start with a privat map instead of aborting. * analyze tool now features a confusion matrix. * pruning panel now more "SNNS-like". You do not need to close the panel anymore before pruning a network. * Changes in batchman -- batchman can now handle DLVQ training -- new batchman command "setActFunc" allows the changing of unit activation functions from within the training script. -- batchman output now with "\#" prefix. This enables direct processing by a lot of unix tools like gnuplot. -- batchman now automatically converts function parameters to correct type instead of aborting. -- jogWeights can now also be called from batchman. -- batchman catches some non-fatal signals (SIGINT, SIGTERM, ...) and sets the internal variable SIGNAL so that the script can react to them. -- batchman features ResetNet function (e.g. for Jordan networks). * new tool "linknets" introduced to combine existing networks. * new tools "td_bignet" and "ff_bignet" introduced for script-based generation of network files; Old tool "bignet" removed. * displays will be refreshed more often when using the graphical editor * weight and projection display with changed color scale. They now match the 2D-display scale. * pat_sel now can handle pattern files with multi-line comments * manpages now available for most of the SNNS programs. * the number of things stored in an xgui configuration file was greatly enhanced. * Extensive debugging: -- batchman computes MSE now correctly from the number of (sub-) patterns. -- RBFs receive now correct number of parameters. -- spurious segmentation faults in the graphical editor tracked and eliminated. -- segmentation fault when training on huge pattern files cleared. -- various seg-faults under single operating systems tracked and cleared. -- netperf now can test on networks that need multiple training parameters. -- segmentaion faults when displaying 3D-Networks cleared. -- correct default values for initialization functions in batchman. -- the call "TestNet()" prohibited further training in batchman. Now everything works as expected. -- segmentation fault in batchman when doing multiple string concats cleared and memory leak in string operations closed. -- the output of the validation error on the shell window was giving wrong values. -- algorithm SCG now respects special units and handles them correctly -- the description of the learning function parameters in section 4.4 is finally ordered alphabetically. ********************************************************************** What is SNNS ? ********************************************************************** SNNS (Stuttgart Neural Network Simulator) is a software simulator for neural networks on Unix workstations developed at the Institute for Parallel and Distributed High Performance Systems (IPVR) at the University of Stuttgart. The goal of the SNNS project is to create an efficient and flexible simulation environment for research on and application of neural nets. The SNNS simulator consists of two main components: 1) simultor kernel written in C 2) graphical user interface under X11R5 or X11R6 The simulator kernel operates on the internal network data structures of the neural nets and performs all operations of learning and recall. It can also be used without the other parts as a C program embedded in custom applications. It supports arbitrary network topologies and, like RCS, supports the concept of sites. SNNS can be extended by the user with user defined activation functions, output functions, site functions and learning procedures, which are written as simple C programs and linked to the simulator kernel. Currently the following network architectures and learning procedures are included: * Backpropagation (BP) for feedforward networks vanilla (online) BP BP with momentum term and flat spot elimination batch BP BP with Chunkwise Updating * Counterpropagation * Quickprop * Backpercolation 1 * RProp * RProp with Weight Decay * Generalized radial basis functions (RBF) * ART1 * ART2 * ARTMAP * Cascade Correlation * Dynamic LVQ * Backpropagation through time (for recurrent networks) * Quickprop through time (for recurrent networks) * Self-organizing maps (Kohonen maps) * TDNN (time-delay networks) with Backpropagation * Jordan networks * Elman networks and extended hierarchical Elman networks * Associative Memory * RBF_DDA * Simulated Annealing * Monte Carlo. * Pruned-Cascade-Correlation A number of network pruning algorithms are available as well: * Optimal Brain Damage (OBD), * Optimal Brain Surgeon (OBS), * Skeletonization, * Magnitute based pruning (Mag). The graphical user interface XGUI (X Graphical User Interface), built on top of the kernel, gives a 2D and a 3D graphical representation of the neural networks and controls the kernel during the simulation run. In addition, the 2D user interface has an integrated network editor which can be used to directly create, manipulate and visualize neural nets in various ways. ********************************************************************** Machine architectures on which SNNSv4.2 is available ********************************************************************** We have tested SNNSv4.2 on the following machines and operating systems: machine type OS user interface with SUN Sparcstation Solaris X11R5, X11R6, OW 3.0 SGI Indigo II IRIX 5.1, 5.2 X11R5 DEC Alpha Workstation OSF1 V4.0 X11R5 IBM RS 6000 AIX V3.2 X11R5 HP 9000/720, 730 HP/UX 9.0.1 X11R5 IBM-PC Pentium Linux X11R6 Our parallel versions of SNNS are only available for research partners with whom we have sponsored joint research projects. These parallel versions include Neurocomputer Adaptive Solutions CNAPS serverII SIMD computer MasPar MP-1, MP-2 MIMD computer Intel Paragon XP/S5, Connection Machine CM-5 ********************************************************************** SNNSv4.2 licensing terms (short) ********************************************************************** SNNSv4.2 is available NOW free of charge for research purposes under a GNU-style copyright agreement. See the license agreement in the user manual and in the file Readme.license of the distribution for details. SNNS is (C) (Copyright) 1990-96 SNNS Group, Institute for Parallel and Distributed High-Performance Systems (IPVR), University of Stuttgart, Breitwiesenstrasse 20-22, 70565 Stuttgart, Germany, and (C) (Copyright) 1996-98 SNNS Group, Wilhelm Schickard Institute for Computer Science, University of Tuebingen, Koestlinstr. 6, 72074 Tuebingen, Germany. SNNSv4.2 can only be obtained by anonymous ftp over the Internet. See the detailed description of how to obtain SNNS below. We don't have the time and capacity to send tapes or floppy disks, so don't ask. SNNSv4.2 is also too large to be mailed by e-mail, so don't ask for that, either. You may, however, obtain the unmodified SNNSv4.2 distribution from other sites which already have obtained it, under the terms of our license agreement, if you are unable to connect to our machine. Note that SNNS has not been tested extensively in different computer environments and is a research tool with frequent substantial changes. It should be obvious that WE DO NOT GUARANTEE ANYTHING. We are also not staffed to answer problems with SNNS or to fix bugs quickly. For questions and/or comments concerning SNNS we refer you to the SNNS mailing list. To subscribe, send a mail to SNNS-Mail-Request@Informatik.Uni-Tuebingen.DE With the one line message (in the mail body, not in the subject) subscribe ********************************************************************** How to obtain SNNSv4.2 ********************************************************************** The SNNS simulator can be obtained via anonymous ftp from host ftp.informatik.uni-tuebingen.de (129.69.211.2) in the subdirectory /pub/SNNS as file SNNSv4.2.tar.Z (2.60 MB) or in gzipped form as file SNNSv4.2.tar.gz (2.18 MB) Be sure to set the ftp mode to binary before transmission of the files. Also watch out for possible higher version numbers, patches or Readme files in the above directory /pub/SNNS . After successful transmission of the file move it to the directory where you want to install SNNS, uncompress and extract the file with the Unix command tar xvfz SNNSv4.2.tar.gz The SNNS distribution includes full source code, installation procedures for supported machine architectures and some simple examples of trained networks. The PostScript version of the user manual can be obtained as file SNNSv4.2.Manual.ps.Z (1.61 MB) or SNNSv4.2.Manual.ps.gz (1.35 MB) or in 15 parts as files SNNSv4.2.Manual.part01.ps.Z ... SNNSv4.2.Manual.part15.ps.Z These parts are all under 1 MB in size when uncompressed and should be printable on any PostScript printer.Again remember to set the ftp mode to binary before transmission of the file(s). There is also an Implementation Manual available as file SNNSv4.2.Implem.ps.Z (0.24 MB) and a set of extension chapters for those users who just printed the manual of version 4.2 in file SNNSv4.2.Manual.Extensions.ps.Z (0.30MB) More information about SNNS as well as a html version of the manual can be found at http://www-ra.informatik.uni-tuebingen.de/SNNS/ A printed version of the 4.2 manual is also available. To cover for the cost of printing and postage you should include DM 20.- for Europe, US $ 20.- overseas in a request for any bound manual; (this covers surface mail postage).