Local Field Distribution in Attractor Neural Networks: Effect of the Stimulus

Abstract

We investigate the distribution of the local fields in attractor neural networks using a simple model similar to that studied by Amit and Brunel (1993). This model consists of two layers of neurons: input layer and recurrent layer. The stimulus of the input layer, by means of feed-forward connections, creates receptive fields on the neurons of the recurrent layer. We discuss the dynamic properties and the basins of attraction of the network and propose a procedure to stabilize the system, driving it into a basin of attraction, without introducing explicitly inhibitory subnetwork. We also investigate the effect of the stimulus applied on the attractor network and propose a criterium of learnability of new patterns. The proposed techniques can be applied to more complex topologies of linked recurrent networks and could be used for designing hardware realizations with application to neurocomputers.