Multi‐Functional Optical Neural Network in a Disordered Medium

Abstract Optical neural networks leverage the speed and parallelism of light to process information. However, many prior works are highly resource‐intensive and lack multi‐task ability. Here, a multi‐functional optical neural network based on wavefront shaping in a nonlinear disordered media is proposed. The setup can serve as either a linear or nonlinear neural network, where the input field is modulated by a spatial light modulator, and all the modulated modes are interconnected via random scattering in the disordered media. An adaptive training pipeline is built based on the separable natural evolutionary strategies. This versatile and low‐cost system can perform a range of computational imaging tasks, including beam mode decomposition and image reconstruction through diffusers, without necessitating wavefront sensing or transmission matrix measurement. Additionally, its applicability to classical machine learning tasks, specifically image classification is demonstrated. These results collectively indicate that this optical neural network presents a promising platform for high‐speed computational imaging and machine learning with potential implications for various fields.