Minimalist Optical Neural Computing: Optical Diffractive Neural Network by 2-level Quantized Pixel-Wise Optical Encoding

Abstract

Diffractive optical neural networks (DONNs) offer high-speed, energy-efficient artificial intelligence (AI) computation but face challenges with optical misalignment and model-to-reality gaps. In this work, an ultra-simplified DONN architecture based on a digital mirror device (DMD) and camera, dubbed as m-DONN, is introduced and experimentally validated. Notably, within the m-DONN framework, the DMD acts as both the input layer and the solitary hidden layer, which is trained with 2-level quantization, markedly differing from the configuration found in traditional DONNs. This minimalism and binarization of the diffraction layer can result in a highly nonlinear correlation between the encoded input information and the output. A 10-classification accuracy of over 82% is achieved on the MNIST dataset in both theoretical modeling and experimental measurements, utilizing over 10 000 test samples. Furthermore, this m-DONN is employed to construct an online reinforcement learning agent capable of dynamically stabilizing a virtual inverted pendulum. The inherent simplicity of the proposed optical computing system, coupled with the cost-effective implementation using either active or passive key optical components, not only demonstrates an extremely powerful yet simple optical neuromorphic setup but also paves the way for the acceleration of optoelectronic AI applications across a variety of scenarios.