In-Sensor Reservoir Computing Based on Self-Rectifying TiOx Photosynapse for Image Recognition and Speech Signal Processing

Abstract

A homojunction memristor with a double layer of resistive function was prepared by the sol–gel method. Electrical measurements show that the F-doped SnO₂ (FTO)/TiO_x/Au memristor exhibits typical volatility, endowing the device with excellent synaptic properties. By analyzing the voltammetry characteristics of the FTO/TiO_x/Au memristor, it is suggested that the titanium oxide homomode memristor exhibits self-rectifying and analog switching characteristics, which can provide inherent advantages for large-scale array integration for high-efficiency bioinspired computing as well as in-sensing artificial vision computing. For its electric switchings, this photoelectronic memristor holds a low operating power consumption, dual-pulse dissimilation learning rules that can faithfully simulate the multifunction of synapses, and learning–forgetting–relearning rules. In terms of optical properties, our memristor exhibits excellent positive photoconductance memory effect (PPM), and the conductance states are sensitive to the light intensity and frequency, significantly presenting synaptic habituation characteristics that are fully controlled. More importantly, integrating the light and electric stimuli control, the conductance weight can be highly linear update, largely accelerating training processing for the reservoir computing (RC). Recognition of image and speech signal processing as the artificial intelligence task is demonstrated using our photoelectronic TiO_x memristor, showing over 90% accuracy after several epoch training. This in-sensor reservoir computing based on self-rectifying TiO_x photosynapse gives a brand-new horizon on neuromorphic vision computing in future edge computing.