Homogeneous reservoir computing system based on CMOS-compatible optoelectronic synapses and memristors.

Dynamic image detection and recognition are critical in machine vision. In traditional visual systems, sensing, storage, and computing are usually separated, leading to significant data transfer overhead. In this work, we presented a homogeneous silicon-based reservoir computing (Si-RC) system, which is composed of SiNx optoelectronic synapses and memristors for the real-time recognition of dynamic images. ITO/SiNx/Pt optoelectronic synapses demonstrated excitatory postsynaptic current (EPSC) and paired-pulse facilitation (PPF) behaviors, functioning as virtual nodes in the reservoir to encode the temporal features of optical image signals. Ti/SiO2/SiNx/Pt memristors showcased uniform resistive switching and electronic synaptic properties, with a coefficient of variation as low as 0.90% in long-term potentiation/depression after applying 104 consecutive voltage pulses. The Ti/SiO2/SiNx/Pt memristors were utilized as trainable weights in the readout layer of Si-RC system, enabling learning and inference. The Si-RC system sensed and preprocessed ultraviolet signals from rocket exhaust plumes then identified the rocket's movement directions, achieving 100% recognition accuracy. Compared to multilayer perceptron network, the system demonstrated a 1.9-fold improvement in training speed.