Bioinspired In-Sensor Reservoir Computing for Self-Adaptive Visual Recognition with Two-Dimensional Dual-Mode Phototransistors

Artificial visual systems that dynamically process spatiotemporal optoelectronic signals under complex real-life environments bear a wide spectrum of edge applications. Despite significant progress in optoelectronic sensors and neuromorphic computing algorithms, developing visual systems that can adapt to a broad illumination range while retaining high performance, high efficiency, and low training costs remains a challenge. Here, this work reports a bioinspired in-sensor reservoir computing (RC) for self-adaptive visual recognition. By leveraging voltage-tunable photoresponses of the MoS2-based phototransistor array, the RC system demonstrates both scotopic and photopic adaptation functions and maintains a recognition accuracy of 91%. The horizontal modulation (HM) block enables the reservoir to adapt automatically in real-time under changing illumination conditions, yielding a 90.64% recognition accuracy (14.21% improvement over conventional RC systems). These results pave the way for the emergence of a reconfigurable in-sensor RC system with broad applications and enhanced performance for an efficient artificial vision system at the edge.