Programmable optoelectronic memristors for energy-efficient adaptive binarized spiking neural networks

Abstract

Hardware realization of brain-inspired vision critically relies on neuromorphic devices that couple sensing, computation, and adaptive learning, yet such integration remains limited by unstable ion dynamics, poor controllability of device plasticity, and the learning-rate sensitivity of binarized networks. Here, we introduce an energy-efficient optoelectronic memristive platform based on Ag/Sb₂Se₃/SnO₂/FTO devices, in which a 10 nm SnO₂ interfacial layer confines Ag-ion migration, enabling stable bipolar resistive switching with an on/off ratio of ∼10 and retention exceeding 10⁴ s. The device exhibits fast polarity-dependent photoresponses with rise and fall times of 722 and 340 μs, respectively, together with a near-linear mapping between conductance and the photoresponse coefficient. Exploiting this property, convolutional kernels are directly encoded at the sensor level for self-powered in-sensor preprocessing and noise suppression. Coupled with a variable-learning-rate binarized spiking neural network, the system achieves 89.8% accuracy on Fashion-MNIST with an inference energy of 2.8 μJ, corresponding to a 94.4% reduction relative to conventional CNNs.