Optimal Dilution and High Capacity Associative Memories Stephen Hunt, Neil Davey and Ray Frank

Associative memories with recurrent connectivity can be built from networks of perceptrons and trained using the perceptron learning rule. Recent work has shown how a perceptron can be diluted, so that incoming connections are removed to produce an optimal perceptron. Such a perceptron classifies its training set with optimal margin and minimal connectivity. Here this technique is applied to perceptrons in associative memory networks. The main result shows that, by using these dilution methods, effective associative memories can be built with very sparse connectivity. Keywords: Associative memory, Hopfield network, optimal dilution. 1