Improving the Variation-Tolerance of Memristor Synapse by Selecting the Optimal Memristance

Abstract

Memristor crossbar array that has a lot of advantageous features such as non-volatile, high-density, and low-power, is potentially used for realizing artificial neural networks. One of the important factors affecting the performance of crossbar circuits is memristance variation. The variation of memristance causes the variation of synaptic weights resulting in the degradation of the performance of the neural networks. In this paper, the variation-tolerance of memristor synapse is improved by minimized the gradient of the synaptic weight function. To reduce the gradient of synaptic weight function, the memristance values of memristors must be close to the high resistance state. This can be achieved by selecting the value of bias resistance because memristance values of memristors in crossbar-based neural network are distributed around the bias resistance value. In the simulation result, we measure the recognition of crossbar circuit for recognizing the MNIST handwritten digits with the variation is in range of 0% to 10% and the value of bias resistance varies from 30KΩ to 80KΩ. For the bias resistance is as low as 30KΩ, the recognition rate of the crossbar circuit is as low as 17%, when the variation is 10%. When the bias resistance is 80KΩ, the recognition is 70%, when the variation is 10%. When the range of memristance is close to the high resistance state, the recognition rate is improved by 53% when the variation is 10%.