Memristor crossbar based low cost classifiers and their applications

Abstract

Existing studies have demonstrated the use of memristor crossbars for learning linearly separable functions. The memristors are used as analog synaptic weights, thus allowing the memristor crossbar to evaluate a large number of multiplication and addition operations concurrently in the analog domain. Non-linearly separable functions can be implemented by cascading two or more crossbars, with each crossbar implementing a linearly separable function. The training circuits for these cascaded crossbars implementing non-linearly separable functions requires more complex logic than for linearly separable functions. In this paper we have implemented non-linear classifiers utilizing multiple linear separators and thus can utilize a simpler training circuit. We have examined the implementation of Boolean functions and motion detection applications as case studies.