Fiber-based diffractive deep neural network

Abstract

Optical computing has reemerged as a promising alternative computing paradigm for providing energy-efficient information processing in the age of artificial intelligence. Among various photonic neural network platforms, diffractive optical processing systems in free space proved to be high-performance computing with high parallelism. Here, we report fiber-based diffractive deep neural networks by optimizing the linear coupling of the waveguide modes. Our approach demonstrated high performance in various machine learning tasks, such as biomedical disease, fashion, and geospatial classification, with a simple readout layer and all-optically. Operating on linear optics, our architecture performs on par with neural networks even in complex datasets where the data cannot be separated using linear operations. These results will enable efficient and scalable diffractive information processing with waveguides for real-life computing, telecommunications, and imaging applications.