The recall performance of multiple associative memories - a comparison of thresholding strategies

Abstract

The associative network presents a practical and efficient means for the storage and retrieval of data. Although the recall abilities of single-layer associative memories have been well researched [3], little investigation of a similar nature has been done on multi-layer associative memories. An earlier theoretical analysis of recall abilities of single-layer associative networks is extended here to provide a more accurate formulation of empirical results and simulation results are presented in support of this analysis. The recall performance of the advanced distributed associative memory (ADAM ) is shown to deliver significantly better recall performance than multi-layer associative memories using Willshaw thresholding at all layers. Intermediate layer parameters that would yield optimum recall performance are investigated and shown to compare favourably with values suggested by information-theoretic derivations.