Analog Computing for Deep Learning: Algorithms, Materials Architectures

Abstract

Analog, or neuromorphic, computing for Deep Learning (DL) utilizes the fact that matrix manipulations that are inherent in the back-propagation algorithm, can be performed at constant time, in parallel, on arrays with nonvolatile memory (NVM) elements in which the weights are encoded. We discuss the NVM material requirements that need to be met to achieve a classification accuracy on par with the conventional digital approaches, discuss advantages and drawbacks, and highlight opportunities that can take advantage using analog arrays.