Preprint

[Gregory Kohring]

The following preprint is currently available.
                                       -- G.A. Kohring




                    Finite-State Neural Networks:
         A Step Towards the Simulation of Very Large Systems

                            G.A. Kohring
                       HLRZ an der KFA Julich
               (Supercomputing Center at the KFA Julich)

                              Abstract

Neural networks composed of neurons with Q_N states and synapses with
Q_Jstates are studied analytically and numerically. Analytically it is
shown that these finite-state networks are up to 25 times more efficient
at information storage than networks with continuous synapses. In order
to take the utmost advantage of networks with finite-state elements, a
multi-neuron and multi-synapse coding scheme is introduced which allows
the simulation of networks having over one billion couplings at a speed
of 7.1 billion coupling evaluations per second on a single processor of
the Cray-YMP. A local learning algorithm is also introduced which allows
for the efficient training of large networks with finite-state elements.

Key Words: Neural Networks, Multi-Spin Coding, Replica Method,
 Finite-State Networks, Learning Algorithms

HLRZ-33/90

Send Correspondence and request for preprints to:

G.A. Kohring
HLRZ an der KFA Julich
Postfach 1913
D-5170 Julich, West Germany