Pyramid-ladder diffractive neural network for visual recognition

When applied to visual tasks, the diffractive neural network (DNN) based on light propagation between diffractive layers has advantages of all-optical, flexible, low-cost, easy to use as a plug-in, and also has good scalability. As the DNN goes deeper, some problems appear: slow convergence, over-fitting of network parameters, disappearance and explosion of gradients. Most of these problems are related to the redundancy of neurons brought by the standard size diffraction layer. Moreover, the energy loss caused by diffraction layer is inevitable. In this paper, we introduced a pyramid-ladder diffractive neural network (PDNN) which is high efficiency, low cost, strong scalability, able to improve DNN performance through imitation of residual structure and restrict gradients without additional optical hardware. What is more, the unique structure makes achievable layers number of PDNN break through the limitations of hardware and materials, so as to deal with complex problems such as object tracking. We also extended the network in multiple dimensions to make full use of the light and leverage the pyramid-ladder module as fundamental bricks. Code and an auxiliary software integrating functions of diffraction, 4−f imaging system, supplementary light parameter selection assistance and pre-trained models for mechanism demonstration are available upon request.