Stochastic dynamics and high capacity associative memories

The addition of noise to the deterministic Hopfield network, trained with one shot Hebbian learning, is known to bring benefits in the elimination of spurious attractors. This paper extends the analysis to learning rules that have a much higher capacity. The relative energy of desired and spurious attractors is reported and the affect of adding noise to the dynamics is empirically investigated. It is found that the addition of noise brings even more benefit in the case of the higher capacity rules.

1 Introduction

In this paper we examine how a variety of high capacity associative memory models respond to noise in the dynamics. The networks are all variations on the standard Hopfield model” differing in the weight matrix that is used to embed the set of training patterns.