Analysis of Correlation Matrix Memory and Partial Match-Implications for Cognitive Psychology

Abstract

This paper describes new work on partial match using Correlation Matrix Memory (CMM), a type of binary associative neural network. It has been proposed that CMM can be used as an inference engine for expert systems, and we suggest that a partial match ability is essential to enable a system to deal with real world problems. Now, an emergent property of CMM is an ability to perform partial match, which may make CMM a better choice of inference engine than other methods that do not have partial match. Given this, the partial match characteristics of CMM have been investigated both analytically and experimentally, and these characteristics are shown to be very desirable. CMM partial match performance is also compared with a standard database indexing method that supports partial match (Multilevel Superimposed Coding), which shows CMM to compare well under certain cirumstances, even with this heavily optimised method. Parallels are drawn with cognitive psychology and human memory.