Dynamic Maps and Attractor Phase-Structures in Randomly Dilute Neural Networks

Abstract

There is currently much interest in attractor neural networks as idealized models of associative memory in the cerebral cortex of the brain [1]. They consist of simple neurons driving one another non-linearly via connecting pathways with mutually interfering characteristics, leading to non-trivial global dynamics. The objective is to train the characteristics of the individual pathways so as to lead to a set of ‘basins’ in which the global dynamic activity is attracted towards that associated with the patterns which are to be memorized and recallable. For useful memory there should be many attractor basins, each with macroscopic overlap with a single memorized pattern.