Alleviating Conductance Nonlinearity via Pulse Shape Designs in TaO x Memristive Synapses

Analog resistive switching behavior in emerging nonvolatile memories has facilitated a wide range of potential applications in neuromorphic computing and cognitive tasks. However, the intrinsic nonlinearity (NL) of conductance update has been proven to be highly unfavorable for the implementation of analog synapse-based hardware neural network (HNN). In this brief, we show that the conductance update characteristics of our Pt/TaO x /Ta memristor will be significantly improved by carefully designing the potentiation and depression pulse shapes. Furthermore, measured conductance update characteristics with different degrees ofNL are applied in the simulation of a face classification task. The results show that both the recognition accuracy and the learning speed of this supervised learning scenario are substantially improved by the proposed optimization approach.