The capacity and attractor basins of associative memory models

The performance characteristics of five variants of the Hopfield network are examined. Two performance metrics are used: memory capacity, and a measure of the size of basins of attraction. We find that the post-training adjustment of processor thresholds has, at best, little or no effect on performance, and at worst a significant negative effect. The adoption of a local learning rule can, however, give rise to significant performance gains.