Bifunctional Photonic Synapse With Short‐Term and Long‐Term Plasticity for Neuromorphic Image Recognition

ABSTRACT Advances in artificial synaptic devices are indispensable for deepening the physical underpinnings of neural networks and for diversifying the tasks that artificial intelligence can tackle. Yet, the photonic synapses demonstrated so far mostly demand electrical readout or intricate heterostructures, and none has offered a built‐in, material‐level mechanism that unites short‐term plasticity with long‐term memory in a monolayer, all‐optical platform. We introduce Ti‐doped CaSb 2 O 6 as a purely photonic synapse whose bifunctional shallow and deep traps natively partition volatile and non‐volatile memory. Dual‐wavelength UV control (275 nm excitation/365 nm inhibition) elicits short‐term facilitation, spike‐number‐dependent potentiation, post‐tetanic potentiation, and erasable storage. From these dynamics we derive an “Opto‐Logistic” activation function and embed it in a lightweight neural network hosted on a microcontroller, demonstrating an AI “dog” that autonomously classifies vegetables and can be retrained for new categories like fruits within a reservoir‐computing framework. The findings reveal how persistent luminescence can mirror biological synaptic physics and furnish both material and system‐level design rules for scalable photonic neuromorphic processors.