Attractors in Associative Memories with Stochastic Weights

Abstract

Neural associative memories are considered in which the elements of the weight matrix are taken to be stochastic variables. The probability density function of each weight is given by the solution of Schrodinger’s diffusion equation. The weights of the proposed associative memories are updated with the use of a learning algorithm that satisfies quantum mechanics postulates. This learning rule is proven to satisfy two basic postulates of quantum mechanics: (a) existence in superimposing states, (b) evolution between the superimposing states with the use of unitary operators. Taking the elements of the weight matrix of the associative memory to be stochastic variables means that the initial weight matrix can be decomposed into a superposition of associative memories. This is equivalent to mapping the fundamental memories (attractors) of the associative memory into the vector spaces which are spanned by the eigenvectors of the superimposing matrices and which are related to each other via unitary rotations. In this way, it can be shown that the storage capacity of the associative memories with stochastic weights increases exponentially with respect to the storage capacity of conventional associative memories.