Performance comparison of correlation matrix memory implementations

Abstract

This paper compares the performance of software and hardware implementations of binary correlation matrix memory (CMM). CMM is a simple, one-layer neural network with a Hebbian learning rule which offers excellent speed and scalability advantages. CMM building blocks form the basis of the AURA neural network system which has been applied to a broad range of practical problems. The paper presents the results of a performance comparison between recent software and hardware implementations of binary CMM. The results show that the hardware implementation provides a best-case speed-up of 50 over the software implementation. Finally, some areas for further improvement in the hardware implementation are identified.