AgPS3-Based Electrical Synapses with Tunable Multilevel Conductance for Noise-Robust Image Classification

Abstract

Developing high-performance artificial synaptic electronics with novel functional materials is both crucial and challenging for the establishment of brain-inspired neuromorphic systems. Neuromorphic computing requires memristive devices with controllable conductance modulation and hardware-compatible information processing capabilities. Herein, layered AgPS3 is innovatively introduced to develop synaptic devices, and its neuromorphic application potential is investigated. Under diverse electrical stimulation protocols, the device exhibits arbitrarily tunable multilevel conductance states, enabling the emulation of biologically synaptic functions of stimulus-dependent weight update, memory transition, and frequency-selective signal processing. In particular, a AgPS3 synapse-based Gaussian front-end for medical image preprocessing is well designed, which significantly enhances the noise robustness of BloodMNIST classification implemented via convolutional neural networks (CNNs). These results demonstrate that AgPS3 can serve as a promising candidate material for artificial synapses, highlighting its great prospects in edge-intelligent sensing, neuromorphic computing, and hardware-efficient image processing.