Abstract

Diffractive neural network is a versatile optical computing platform for accelerating deep learning inference. Whereas enabling the free-space optics for the multi-task processing capability requires complicated and costly system reconfigurations. Herein, we turn the neural network driven holography to a holographic classifier, where multiple image recognition tasks can be tackled simultaneously.

It results in an end-to-end optoelectronic neural network, which processes information from the target image (input layer), through the generated phase map (hidden layer), to the resulting optical pattern (output layer). The simple optical architecture and flexible neural holographic algorithm facilitate the accurate model-to-reality implementation.

The proposed system has potential applications for high-speed information processing, and intelligent control of beam splitting and steering.