A two-pathway informon theory of conditioning and adaptive pattern recognition

Abstract

A neural network theory is proposed which offers an explanation of many of the facts of classical and operant conditioning and adaptive pattern recognition. Interconnected networks of units have been studied and simulated which embody only two rules; firstly, units have inputs from pathways of variable and of fixed conductivity; secondly, the conductivity of a variable pathway is made proportional to the negative of the mutual information function between the signals at its input and output. The signal in a fixed pathway indicates whether the total input to the variable pathways is a member or not of some class. After a learning phase in which the unit, called an informon, receives such labelled inputs, it is able to predict the class of future unlabelled inputs. Such units are stable and their steady state can be calculated.