A new stochastic mutiplier for deep neural networks

Abstract

An XNOR gate is the most commonly used multiplier in bipolar encoded stochastic deep neural networks, but it is not suitable due to the inaccuracy in processing near-zero values. In this paper, we introduce a novel circuit that multiplies near-zero values more accurately and assess its performance with MNIST and CIFAR-10. For the CIFAR-10 dataset, the use of the proposed multipliers gives accuracy of 60.59%, improving by 11.64%p over the XNOR multiplier implementation.