In‐Sensor Reservoir Computing Employing Reconfigurable Optoelectronic Transistors for Multi‐Task Learning

ABSTRACT The edge deployment of artificial intelligence has driven the exploitation of compact, energy‐efficient information processing systems that integrate sensing, memory, and multi‐task processing functions. However, conventional vision systems suffer from significant energy/time overhead, extra hardware costs, and an unaffordable algorithm. Herein, we demonstrate an in‐sensor computing system employing reconfigurable optoelectronic transistors (ROETs) for multi‐task learning. These transistors exhibit reconfigurable volatile and nonvolatile characteristics under both optical and electrical stimuli. Capitalizing on this reconfigurability, we establish an in‐sensor reservoir computing (RC) system operating in multi‐signal modes: volatile dynamics function as the reservoir, whereas nonvolatile properties configure the readout layer. The abundant optoelectronic reservoir states display exceptional feature separability and prolonged stability in the ambient atmosphere. Such a reliable RC system successfully achieves multi‐task processing of images. Notably, under the optoelectronic coordination mode, it effectively alleviates feature degradation while sustaining consistently high recognition accuracy. Furthermore, the system exhibits remarkable dynamic information processing capabilities, achieving recognition accuracies of 89.02% for dynamic gestures and 96.04% for moving vehicles recognition, respectively. Supplemental functionalities, including light adaptation and image sharpening, are also implemented. This work presents a configurable multimodal platform featuring a flexible in‐sensor reservoir computing architecture, providing a potential solution for efficient multi‐task processing.