Metasurface-enabled diffractive neural networks for multi-label classification

Optical neural networks have attracted considerable attention due to their capability to facilitate high-speed and low-power neuromorphic computation at the physical level. Diffractive neural networks represent a class of optical neural network architectures that are based on the Huygens-Fresnel principle. This paper presents a diffractive neural network based on silicon-based dielectric metasurfaces. A multi-label classification task, which improves classification efficiency and increases the ability of computation compared to traditional handwritten digit classification tasks, is proposed as a means of training this neural network and exploring its classification capabilities. The proposed approach offers a new way for the application of diffractive neural networks in the domain of target classification and recognition.