A RRAM based Max-Pooling Scheme for Convolutional Neural Network

In this paper, a novel max-pooling scheme for the convolutional neural network (CNN) is proposed and implemented by a simple analog RRAM based circuit. By taking advantage of the RRAM multilevel characteristics in the current driven mode, the proposed scheme can automatically detect and memorize the largest accumulative current in the crossbar during pooling operation. The simulation result of MNIST is studied and proved that this scheme can save 90.0% energy and increase 51.5% in speed. The impacts of typical non-ideal effects on accuracy are also investigated.