Research On Design And Training Method Of Optical Scattering Neural Network

In recent years,with the rapid development of neural networks,a variety of neural network frameworks have been proposed,showing strong application potential in image recognition and speech processing.However,as the neural network structure becomes growing complex,the requirements of computing power are also higher and higher.In terms of integrated circuits,due to the gradual failure of Moore's Law,modern transistors are increasingly unable to meet the computing power requirements of artificial intelligence(AI)technology.In addition,electronic signals have problems such as susceptibility to interference,high energy consumption,and high heat dissipation risks.As an information carrier,photons have the advantages of high energy efficiency and high-speed large-scale parallel processing capabilities.Recently,the rapid development of optical devices,such as optical matrix vector multiplier,optical interconnection technology,etc.,has successfully solved the realization of optical matrix multiplication and optical nonlinearity.In addition,a variety of new optical neural network architectures and new methods have been proposed,such as photoelectric hybrid neural network,all-optical neural network,etc.,which have greatly promoted the development of optical intelligence.Based on the above results and factors,the main research aspects of this article are as follows:1.First,the semi-analytical method based on Maxwell's equations is used to describe the physical properties of the bidirectional scattering propagation of the optical scattering medium,the physical properties of the scattering matrix are described by the matrix method,and the optical scattering neural network is constructed by using the multilayer scattering medium.The layers extend to multiple neural layers.2.Design a new type of optical scattering neural network framework based on scattering medium imaging.According to the bidirectional scattering physical model,the corresponding bidirectional digital complex-valued neural network is constructed.According to the forward propagation network of the bidirectional digital complexvalued neural network,the complex-valued backward propagation of the digital weight update is guided by the complex-valued derivative chain rule and the gradient descent method algorithm.3.A matrix method is used to describe the bidirectional scattering imaging model of the optical scattering medium,and the weight parameters of the digital complex neural network are realized by adjusting the physical parameters of the optical scattering medium.4.Use multi-layer stacked optical scattering media to construct a single-layer intelligent unit of optical scattering neural network,cascade single-layer intelligent units to construct a deep physical model of multilayer scattering network,and use nonlinear photoelectric material film to achieve nonlinear activation to complete optical scattering neural network Finally,the effectiveness of the intelligent classification of the optical scattering neural network is verified on the MNIST data set.