Advances and progress of diffractive deep neural networks

Abstract

In recent years, artificial intelligence has achieved unprecedented development, and deep learning, represented by neural networks, plays an important role. After the emergence of large-scale pre-trained models with trillions of parameters, the model performance is significantly improved while the burden of computational resources and energy consumption of hardware devices are also increased simultaneously, thus limiting its application in more practical scenarios. Compared with neural networks implemented based on electronic devices, those implemented based on optical devices are called optical neural networks, which have unique properties to overcome the dilemma above. One of the most representative works of optical neural networks these years is the diffractive deep neural network (D²NN). In this paper, the research progress of D²NNs is summarized in four aspects: basic theory, further analysis, improvement, and application. Besides, it is analyzed that the common defect of D²NNs from simulation to physical fabrication, and corresponding theoretical improvement method is also proposed. Meanwhile, to further reduce the impact due to the gap between simulation and physical implementation, and to enhance the robustness of the model, the D²NN training method based on generative adversarial network (GAN) is proposed. The D²NN combines optical transmission with deep learning to achieve complex pattern recognition tasks in the optical domain at the speed of light. It is believed that under continuous research, the D²NN can play a greater role in optical communications and other fields.