The bidirectional optimization of STDP conductance update characteristics for neural computing

Abstract

Unsupervised Spiking Neural Networks (SNNs) that operate based on the Spike-Timing-Dependent Plasticity (STDP) learning rule have high biological plausibility and are considered the next generation of artificial neural networks. The development of artificial synapse devices with excellent STDP conductance update characteristics is the foundation for achieving highperformance SNN networks. This work constructs a TaOx/TiOx memristive material featuring three distinct oxygen-vacancy concentration regimes, achieving optimized bidirectional STDP conductance updates (LTP and LTD processes) in memristive synapses. Compared with the singlelayer TaOx-based memristor synapse, the LTP part of the STDP performance of the double-layer TaOx/TiOx device has an increased conductance range by 110%, while the LTD part has an increased conductance range by 61%. The factor ratio of the forward and reverse conductance ranges, A+/A-, is closer to 1. Analysis shows that the slower forgetting rate of the tri-level oxygen defects concentration profile in TaOx/TiOx based memristive synapse is the main reason for the optimized STDP performance. The simulation results show that the optimized STDP characteristics can increase the network recognition rate. This paper presents a device structure and process that can effectively regulate the bidirectional conductance update characteristics of STDP in oxide based memristor, which is conducive to promoting the development of high-performance memristor-based neural morphological devices.