Effect of pruning small weights associative memories

In this paper we examine the effect of pruning small weights in correlation associative memories. The effect is investigated by considering a fully interconnected matrix and removing weights satisfying \w(ij)\ less than or equal to epsilon. We show analytically that under suitable constraints on epsilon, the capacity of a sparsely connected associative memory is comparable to that of a fully connected one, without compromising the basins of attraction around the prototype fixed points. Simulation results supporting the analysis are also presented.