Better Performance of Memristive Convolutional Neural Network Due to Stochastic Memristors

Convolutional Neural Network (CNN) has gotten admirable performance in the domain of image recognition. Nevertheless, the training of CNN on CPU or GPU is energy-intensive and time-consuming. Memristor crossbar is an alternative of the specific chip for CNN application. But it is hard to tune the memristor to certain conductance precisely. This work simulates the performance change of memristor-based CNN when memristor is with stochasticity. The simulation results demonstrate that stochastic memristor-based CNN performs better on CIFAR-10 dataset when memristive stochasticity is low. This is an encouragement for the engineer of memristor crossbar chip and edge computing application.