Bio-inspired optoelectronic neuromorphic device based on 2D vdW ferroelectric heterostructure for nonlinearly preprocessing visual information and convolutional operation

Abstract

The human visual system provides important inspiration for designing energy-efficient and sophisticated artificial visual systems. However, integrating nonlinear preprocessing visual information and convolutional operations analogous to those of human in a single device is still in its infancy. In this work, a three-terminal 2D ferroelectric heterostructure consisting of ?-In2Se3/WSe2 is proposed for designing optoelectronic neuromorphic device. In contrast to conventional ferroelectric materials, the narrow bandgap of the ferroelectric ?-In2Se3 enables the device to perceive visible light directly. Nonlinearly preprocessing is adopted by bipolar cells in the retina and computer algorithms. In the device, similar function is achieved by modulating the energy band based on ferroelectricity. The results demonstrate the ability of the device to suppress noise, and the image recognition accuracy is increased from 75% to 92%. Convolutional neural networks play an important role to extract and compress the image information for human to respond to external environment in real time. Based on the unique coupling of ferroelectricity in ?-In2Se3, the convolutional operation is imitated, thus allowing for reduction in image recognition time by 87%. The results provide a promising strategy to integrate diverse bio-inspired neuromorphic behaviors in a single device for artificial intelligence to process high-throughput visual information.