A novel memristor-based restricted Boltzmann machine for contrastive divergence

Abstract

In this letter, we present a novel memristor-based restricted Boltzmann machine (RBM) system for training the brain-scale neural network applications. The proposed system delicately integrates the storage component of neuron outputs and the component of multiply-accumulate (MAC) in memory, allowed operating both of them in the same stage cycle and less memory access for the contrastive divergence (CD) training. Experimental results show that the proposed system delivers significantly 2770x speedup and less than 1% accuracy loss against the x86-CPU platform on RBM applications. On average, it achieves 2.3x faster performance and 2.1x better energy efficiency over recent state-of-the-art RBM training systems.