Harnessing intrinsic memristor randomness with Bayesian neural networks

Abstract

Memristors could be a key driver in the development of new ultra-low energy edge neural network hardware. However, the technology has one major drawback – memristor properties are inherently random. Information cannot be programmed in a precise manner. As a result, efforts to exploit the technology often result in neural network models that are less performant than their software counterparts or require mitigation techniques that can negate potential energy benefits. In this paper we summarise how, alternatively, these intrinsic device properties, previously regarded as non-idealities to be mitigated, are well suited for an alternative approach - Bayesian machine learning. Like resistive memory device properties, Bayesian parameters are described by distributions of probability - offering a more natural pairing of device and algorithm.