Using Correlation Matrix Memories for Inferencing in Expert Systems

This paper presents an analysis of Correlation Matrix Memory (CMM), a binary associative neural network. CMM has been proposed as an inference engine, which could be utilized in an expert system (Austin, 1994). In this role, CMM encodes Tensor Products (Smolensky, 1990) to support Dynamic Variable Binding. The ability to perform partial match efficiently, with a large body of stored information is important for a useful expert system. This paper extends the work on the storage v. error characteristics of CMM and presents new work on the partial match ability of the memory. A comparison is made with an efficient partial match method used in databases, Multi-level Superimposed Coding (Sacks-Davis & Ramamohanarao, 1983), and the partial match performance of CMM is shown to compare well. We conclude that CMM is well suited to its proposed use as an inference engine.