Resampling the optical transmission matrix to accelerate the training of the diffractive deep neural network

• DOI: 10.1364/ao.529516
• Publication Date: 2024-06-18T19:00:12Z
• Authors: Lin Wu; Rui Xia; Ming Zhao; Xi Xiao; Zhixue He; Yongquan Zeng; Zhenyu Yang; Jin Tao
• Keywords: Resampling

The diffractive deep neural network (D 2 NN) enables all-optical implementation of machine learning tasks. During the training, Rayleigh–Sommerfeld (RS) diffraction integral is employed for connecting neurons between neighboring hidden layers. RS formula can be rewritten as a transmission matrix (TM), which allows parallel computation multiple vectorized light fields. However, TM has large size, demanding substantial computational resources, and resulting in long training time. In this paper, we propose to resample free space based on propagation invariant modes (PIMs), thereby reducing size matrix, accelerating simulations. This method large-scale D NN with reduced memory requirements fast speed.