Research And Design Of Photonic Chip Based On Small Fully Connected Neural Network

With the rapid development of technologies such as 5th generation mobile networks(5G),artificial intelligence(AI),and big data and the exponential growing of the amount of data,traditional electronic chips are gradually unable to support the massive amount of data in the calculation due to constraints such as integration and power consumption.Due to the characteristics of high bandwidth,low power consumption,high parallelism,and easy integration,photonic chips can break through the limitations of "electronic bottlenecks" and realize rapid data processing.In this paper,a photoelectric hybrid neural network based on Mach-Zehnder interferometer(MZI)modulator is built with the goal of neural network operation in the optical domain.Through the analysis of the error of the photonic devices in the network,an all-optical neural network chip based on error compensation scheme is designed.In this photonic chip,linear and nonlinear operations of data are both realized in the optical domain,with high precision and large noise tolerance.The main research work of this article is as follows:(1)Using Python script to assist Interconnect simulation software,a photoelectric hybrid neural network is built.Among them,the linear matrix operation is realized in the optical domain through the MZI structure,and the nonlinear activation function algorithm is performed in the electrical domain.Using the iris classification task as the top-level application,the accuracy of the network is 92.38%.(2)Aiming at the problem of network performance degradation caused by devices errors in the photoelectric hybrid neural network,an error compensation scheme is proposed.Firstly,for the error caused by beam splitters in MZI,the structure of adding redundant MZI is adopted,and a method for accurately calculating the specific parameters in dual MZI is proposed.After training the above neural network,it is found that the method can effectively reduce the error of neural network and improve the operation performance.Secondly,for the error of the phase shifter in MZI,the MZI structure with a redundant phase shifter and corresponding adjustment algorithm are proposed.Research shows that this structure also can improve the accuracy of the neural network and increase the stability of the network.(3)After the error compensation scheme,a structure with redundant MZI and redundant phase shifter is obtained.Through the self-configuration method,the structure can resist the network performance degradation caused by the beam splitters and the phase shifters errors,can perform linear and nonlinear operations at the same time.Based on this,the design scheme of photonic chip based on neural network is realized.After training this neural network,the iris classification task can reach 96.40% accuracy.Comparing its performance with the photoelectric hybrid neural network,the research shows that the all-optical neural network has higher accuracy and higher noise immunity.