CENTER FOR COMPUTATIONAL MATHEMATICS COLLOQUIUM
                  UNIVERSITY OF COLORADO AT DENVER


TITLE:   INTRODUCTION TO REFRACTORY NEURAL NETWORKS

SPEAKER: Sergey K. Aityan, Auto-Trol Technology, Denver, aityan@auto-trol.com

DATE:    Wednesday, October 5, 1994

PLACE:   Math Conference Room - Suite 540
         UCD Building, 1250 14th St., Denver

TIME:    2:30 pm (Refreshments served at 2:15 pm)

ABSTRACT: Conventional artificial neural networks are based upon McCulloch and 
Pitts neural model which takes only active and rest states of the neuron into 
account. The active state of the McCulloch and Pitts neuron can be lasting as 
long as a superthreshold input is being hold to the neuron. Although this model 
was a very important in developing the approach of artificial neural networks, 
the model does not accounts for the refractory state of the biological neuron 
that restricts thebiological relevancy of the McCulloch and Pitts model.

The goal of the present paper is to introduce a biologically relevant model of 
the refractory neuron and to demonstrate the advantages of recurrent refractory 
neural networks which represent a biologically relevant approach in neural 
modeling.

A concept of refractory neural networks is presented. The refractoriness is 
considered an important feature of neural excitation that has significant 
impact on the paradigm of information processing by artificial neural networks. 
The model refractory neural networks are based on a generalization of the 
McCulloch and Pitts neural model by addition of a refractory state of the 
neuron. The principal novelty in the approach of the refractory neural networks 
consists of a dependence of the neuron response function on the phase of the 
neuron firing cycle. The information processing paradigm for recurrent 
refractory neural networks (RRNN) is based upon dynamical vector-time pattern 
rather than static vectors.

A simple three-neuron RNNN is shown to be able to perform all binary logical 
functions, including XOR. The abilities of the refractory neural networks can 
be explained in terms of dynamical firing vector-time pattern rather than the 
static pattern mapping performed by most conventional neural networks. A 
training mechanism for RRNN implements dynamical vector-time mapping adjustment.