Ex-situ training of large memristor crossbars for neural network applications

摘要

Memristor crossbar arrays carry out multiply-add operations in parallel in the analog domain, and can enable neuromorphic systems with high throughput at low energy and area consumption. Neural networks need to be trained prior to use. This work considers ex-situ training where the weights pre-trained by a software implementation are then programmed into the hardware. Existing ex-situ training approaches for memristor crossbars do not consider sneak path currents, and they may work only for neural networks implemented using small crossbars. Ex-situ training in large crossbars, without considering sneak paths, reduces the application recognition accuracy significantly due to the increased number of sneak current paths. This paper proposes ex-situ training approaches for both 0T1M and 1T1M crossbars that account for crossbar sneak paths and the stochasticity inherent in memristor switching. To carry out the simulation of these training approaches, a framework for fast and accurate simulation of large memristor crossbars was developed. The results in this work show that 0T1M crossbar based systems can be 17-83% smaller in area than 1T1M crossbar based systems.