Sign constrained high capacity associative memory models.

Abstract

Biological neural networks do not allow the synapses to choose their own sign: excitatory or inhibitory. The consequences of imposing such a sign-constraint on the weights of the standard Hopfield associative memory architecture, trained using perceptron like learning, are examined in this paper. The capacity and attractor performance of these networks is empirically investigated, with sign-constraints of varying correlation and training sets of varying correlation. It is found that the specific correlation of the signs affects both the capacity and attractor performance in a significant way.