Multimodal Photoelectric Synapses Based on 2D WS ₂ /PdSe ₂ Heterostructures for High‐Accuracy Neuromorphic Vision

ABSTRACT Neuromorphic vision computing offers a promising solution to machine vision's arithmetic bottleneck. Single devices integrating perception, processing, and memory functions have attracted considerable interest, though maintaining effective hardware‐software coordination remains challenging. In this work, we demonstrate a neuromorphic vision systems (NVS) incorporating photoelectrically modulation neuromorphic devices with in‐device computing capabilities. The system employs a three‐terminal hardware configuration utilizing a 2D WS 2 /PdSe 2 heterostructure, which exhibits a high on/off ratio of ∼10 6 and minimal power consumption of 2.4 pJ per event. It demonstrates multimodal analog synaptic behaviors under electrical modulation, including robust synaptic plasticity and weight updateability. Notably, under photoelectric modulation, the device not only enables multi‐parameter tuning of synaptic behaviors across a broad spectral range (460, 532, and 660 nm) via optical pulses, but also can realize bidirectional weight update (LTP/LTD) modulation with wide spectral ranges through electrical pulses. Based on this, we developed a multi‐channel attention residual network (MAResNet) for neuromorphic computing, which achieves 92.8% recognition accuracy on the CIFAR‐10 dataset with an average AUC exceeding 0.98. Moreover, the model exhibits interpretable spatial responses and channel selectivity, further verifying the effectiveness and robustness of the neuromorphic computing framework inspired by biological vision. This work paves the way toward high‐accuracy NVS.