Training, Programming, and Correction Techniques of Memristor-Crossbar Neural Networks with Non-Ideal Effects such as Defects, Variation, and Parasitic Resistance

Abstract

Memristor crossbars can be useful in realizing high-performance and low-power computing hardware especially for realizing edge-intelligence. Unfortunately, however, they have non-ideal effects such as memristor defects, conductance variation, parasitic resistance, etc. For compensating these non-ideal effects, various techniques should be used in implementing neural networks with memristor crossbars. More specifically, the memristor defects should be considered in the training process using defect map. The variation in programmed memristor’s conductance can be suppressed using the fine programming method of memristors. Moreover, to reduce errors of crossbar’s currents and voltages due to parasitic resistance, the correction circuit can be added to the crossbar peripheral. In this paper, these techniques are explained and verified to be able to minimize the recognition rate loss due to the non-ideal effects in the memristor crossbar.