XNOR-BSNN: In-Memory Computing Model for Deep Binarized Spiking Neural Network

This paper proposes a residual binarized spiking neural network (B-SNN) model suited for inmemory computing (IMC) implementation. While in most of the prior arts, due to the nature of spike represented unipolar format, the B-SNN were implemented using either complex or non-regular logic that is not suited for IMC and/or makes the network inflexible. In this work, we present a B-SNN model that permits the direct adoption of a unipolar format spike on the XNOR array, i.e., allows fully exploiting IMC's potential benefit based on the highly regular and simple array structure. Also, instead of indirectly taking the B-SNN model from the trained BNN, we propose a residual model for deep B-SNN networks. The system simulation shows that our trained network achieves reasonably good accuracy (59.11%) on CIFAR100 with very low inference latency (only 8 time-steps BSNN).