High performance associative memory models and weight dilution

Abstract

The consequences of diluting the weights of the standard Hopfield architecture associative memory model, trained using perceptron like learning rules, is examined. A proportion of the weights of the network are removed; this can be done in a symmetric and asymmetric way and both methods are investigated. This paper reports experimental investigations into the consequences of dilution in terms of: capacity, training times and size of basins of attraction. It is concluded that these networks maintain a reasonable performance at fairly high dilution rates.