Algorithmic Optimisation for Memristive Deep Learning Accelerators

Abstract

Conventional computer hardware based on digital (Boolean) logic and the von Neumann architecture, which separates computing and memory, results in large power and time costs in data-intensive applications like deep learning. Memristive-crossbar-based accelerators promise to improve power efficiency and speed by orders of magnitude but suffer from nonidealities, which cause errors. Here, we overview a number of algorithmic approaches that aim to improve the accuracy and robustness of networks implemented on memristive crossbar arrays. Algorithmic optimisation is attractive because it is relatively technology-agnostic and offers many possible options: from improvements of the training procedure to non-disruptive changes at the circuit level.