Strategies to Improve the Accuracy of Memristor-Based Convolutional Neural Networks

Abstract

In this article, we quantify several nonideal characteristics of memristor synaptic devices, such as the limited conductance states, write nonlinearities, and variations, and comprehensively investigate their effects on the convolutional neural network (CNN) performance. Our result shows that the available conductance states ( ${N}_{\text {state}}$ ), asymmetric write nonlinearities, and cycle-to-cycle (C2C) variation are critical factors to the learning accuracy, while symmetric write nonlinearities and device-to-device variation go trivial. We accordingly propose three strategies to mitigate their impacts on CNN performance: 1) limiting the weight range to improve the utilization of ${N}_{\text {state}}$ ; 2) adopting a new “with-read” update scheme to mitigate the effects of asymmetric write nonlinearities; and 3) employing multiple memristors for each kernel element to alleviate the impact of C2C variation. Our work would provide guidance for the hardware implementation and optimization of CNN in memristor crossbar.