Previous: Sesame
Up: NEURAL NETWORK SIMULATION TOOLS
Next: NAME
Previous Page: Sesame
Next Page: NAME

SNNS

General Information
Name
SNNS (Stuttgart Neural Network Simulator) Version 2.0
Source
Dr. Andreas Zell, Universität Stuttgart, Institute for Parallel and Distributed High Performance Systems
Price
Not public domain but `free software', distributed in a licensing agreement free of charge.
Target usage
education, experimentation, research; advanced beginners, experts

Model

Network Architecture
The units are distinguished by the values of several attributes (number, name, io-type, activation, output, bias, activation function, output function, position, subnet number, etc.). The connections to a unit can be grouped by sites and treated differently.
Paradigms
Five network update-modes are implemented. One training method used is online Backpropagation, also known as stochastic gradient method; two versions are available, one without and one with momentum. Furthermore, Quickprop, Counterpropagation, and Backpercolation are implemented. A few examples are included.
Limitations
The net executes in update-steps synchronized by a discrete clock.

Implementation

Display
Visualization is by means of XGUI, it has a manager window containing all elements needed for working with the interface, several displays to show the network in two dimensions, a 3D-view panel, a panel for simulator operations, a special purpose window for the creation of big regular feedforward-nets, and several windows to display help information. Additionally, there are several popup windows. A graphical network editor provides means to generate or modify networks in various ways.
User Interface
Medium sized nets (up to 10.000 units) can be created by XGUI, a high level network description language Nessus for the generation of large nets (10.000 to 1.000.000 units) is available.
Execution
Interactive and batch exexcution is possible. Networks can be saved and loaded, as well as activation values and displays with all setup parameters.
Data Format
The format of the file interface to the kernel is described in the appendix of the user manual. Files of this format may be generated by any other program.
Evaluation
The sum of the squared error over output units and the average error per output unit are reported.
Performance
information not available

System Requirements

Hardware
SUN 3/50, 3/60; SUN SparcSt. SLC, ELC; SUN SparcSt. 2 GX, GS; DECstation 2100, 3100; DECstation 5000/200; IBM-PC 80386; IBM RS 6000/320, 320H; HP 9000/845; HP 9000/720, 730.
Operating System
SunOS 4.1.1, Ultrix V4.2, Interactive Unix, AIX V3.1, HP/UX
Display
monochrome as well as color, X-Windows X11R4 and X11R5, Sun OpenWindows 3.0
Language
C

Installation and Documentation

Installation
Procedures for supported machine architectures are included.
Sources available
yes
Implementation language
C
Documentation
On-line help available, examples and user manual (Report No. 3/92) are included in the distribution.
Support
None, but users are encouraged to send back changes and improvements. Distribution of modified copies is not allowed. There is no warranty of any kind.
User Groups
Mailing list snns@informatik.uni-stuttgart.de



Previous: Sesame
Up: NEURAL NETWORK SIMULATION TOOLS
Next: NAME
Previous Page: Sesame
Next Page: NAME

franz@neuro.informatik.uni-ulm.de
Wed Jan 19 23:46:24 MET 1994