Multi-functional synaptic memristor for neuromorphic pattern recognition and image compression

Abstract

A two-terminal artificial synaptic memristor capable of emulating the discrimination ability in human brain is an essential prerequisite for realizing neuromorphic computing architectures through straightforward crossbar array, however, it is still a challenge yet. Here, a multi-functional synaptic memristor is reported, based on bismuth oxybromide (BiOBr) nanosheets, in which enables advanced pattern-discriminating and image compression synaptic functionality. The device exhibits stable resistance switching with an On/Off ratio of ∼30.4 and pronounced electrically-induced synaptic plasticity. The device array can achieve a classification recognition accuracy of 70.98 % on CIFAR-10 dataset, significantly outperforming the 36.35 % accuracy obtained using traditional gradient descent algorithms. By encoding image pixel values into temporal pulse sequences, the device can enable high-precision image compression, maintaining 94.01 % classification accuracy on MNIST dataset with greatly reduced trainable parameters (from 13550 to 2630) and shortened training time (from 252 to 65 s). These findings suggest BiOBr nanosheets could facilitate efficient memristor-based artificial intelligence applications.