Physical Reservoir Computing with Optoelectronic Memristors for Event-Camera

Abstract The asynchronous and sparse event streams output by event cameras pose a significant challenge to traditional computing architectures, whereas physical reservoir computing (RC) systems based on optoelectronic memristors offer an effective solution for realizing efficient in-sensor computing vision architectures. In this work, we propose and fabricate a Pt/Ag/ZnO/Pt/Ti optoelectronic memristor with a vertically stacked structure. Under ultraviolet light stimulation, the device exhibits excellent photoresponsive sensitivity, short-term synaptic plasticity, and dynamic relaxation characteristics. Upon the application of temporal optical pulses, the device generates highly distinguishable branching evolutionary trajectories, successfully mapping low-dimensional optical signals nonlinearly into a high-dimensional feature space. Leveraging the device’s exceptional spatiotemporal mapping capabilities, we constructed a synergistic physical RC system of hardware and software and applied it to an event-camera data set. The results demonstrate that the system can efficiently decouple highly complex asynchronous spatiotemporal event streams, ultimately achieving a dynamic gesture recognition accuracy of up to 94.17%. This study not only confirms the exceptional mapping capabilities of optoelectronic memristors in processing complex asynchronous spatiotemporal data but also lays a solid hardware foundation for the development of next-generation, highly efficient, and low-power event-driven neuromorphic vision systems.