Optimization of non-linear conductance modulation based on metal oxide memristors

Abstract

Abstract As memristor-simulating synaptic devices have become available in recent years, the optimization on non-linearity degree (NL, related to adjacent conductance values) is unignorable in the promotion of the learning accuracy of systems. Importantly, based on the theoretical support of the Mott theory and the three partial differential equations, and the model of conductive filaments (CFs), we analyzed and summarized the optimization schemes on the physical structure and the extra stimulus signal from the internal factor and external influence, two aspects, respectively. It is worth noting that we divided the extra stimulus signals into two categories, the combined pulse signal and the feedback pulse signal. The former has an internal logical optimized phenomenon, and the composition of only two parts in each cycle leads to a simple peripheral circuit. The latter can obtain an almost linear NL curve in software stimulation because of its feature in real-time adjustment of signals, but it is complex in hardware implementation. In consideration of space and energy consumption, achieving memristor with different resistive switching (RS) layers can be another optimization scheme. Special attention should be paid to the weaker NL, which could improve learning accuracy at the system level only when the value of other non-ideal properties such as the on/off ratio is within a certain range.