Fully Light-Controlled Memory and Neuromorphic Computation in Layered Black Phosphorus

Abstract

Imprinting vision as memory is a core attribute of human cognitive learning. Fundamental to artificial intelligence systems are bioinspired neuromorphic vision components for the visible and invisible segments of the electromagnetic spectrum. Realization of a single imaging unit with a combination of in-built memory and signal processing capability is imperative to deploy efficient brain-like vision systems. However, the lack of a platform that can be fully controlled by light without the need to apply alternating polarity electric signals has hampered this technological advance. Here, a neuromorphic imaging element based on a fully light-modulated 2D semiconductor in a simple reconfigurable phototransistor structure is presented. This standalone device exhibits inherent characteristics that enable neuromorphic image pre-processing and recognition. Fundamentally, the unique photoresponse induced by oxidation-related defects in 2D black phosphorus (BP) is exploited to achieve visual memory, wavelength-selective multibit programming, and erasing functions, which allow in-pixel image pre-processing. Furthermore, all-optically driven neuromorphic computation is demonstrated by machine learning to classify numbers and recognize images with an accuracy of over 90%. The devices provide a promising approach toward neurorobotics, human-machine interaction technologies, and scalable bionic systems with visual data storage/buffering and processing.

Keywords: artificial neural networks; black phosphorus; machine learning; neuromorphics; optical memory.