A comparative analysis of high performance associative memory models.

Abstract

Three variants of the Hopfield network are examined, each of which is trained using a different iterative approximation of the pseudo-inverse rule. All three variants are known to have significantly higher memory capacity than the standard Hopfield model. The performance measure employed in this study is the average size of attractor basins. The three models are tested with both biased and unbiased random data and under different memory loads. We find that a model employing an iterative local learning regime, based upon the least mean squares learning rule, gives the best attractor performance, albeit at the expense of increased training times.