Influence of unique behaviors in an atomic switch operation on hardware-based deep learning

Abstract

Hardware-based deep learning using neuromorphic elements are gathering much attention to substitute the standard von Neuman computational architectures. Atomic switches can be candidate for the operating elements due to their analog resistance change in nonlinear and non-volatile manner. However, there are also several concerns in using atomic switches, such as inaccuracies in resistance control and autonomous weight decay. These characteristics can cause unintentional changes of weights during the learning process. In this study, we simulated how these characteristics of atomic switches influence the accuracy and the power consumption of the deep leaning. By implementing the weight decay, the accuracy remained high despite of the high error level. Power consumption also improved with weight decay in high error level.