Research On Optical Neural Network For Feature Recognition Based On Silicon-based Waveguide

In recent years,artificial neural networks have made breakthrough progress in many areas such as image recognition,pattern recognition,machine translation,and unmanned driving,pushing human society into the intelligent era.The era of big data provides massive data for neural networks to learn,which also poses more severe challenges to the computing power of computers and the application of algorithms.The emergence of artificial intelligence chips has greatly improved computing efficiency after 2015 and has become the hardware foundation to promote the development of artificial intelligence.However,the calculation time and power consumption are still strictly restricted by the system power and bandwidth.Computing speed and power consumption have become one of the difficult problems to be solvedIn recent years,many researches on implementing neural networks based on photonic integrated chips have emerged.Optical signals have the ability to perform complex operations under passive conditions and have the potential to achieve ultra-high-speed parallel operations,providing a theoretical basis for photonic neural networks.Based on silicon-based optical waveguides,this paper proposes a confirmatory study of digital recognition with optical neural networks constructed by Mach-Zehnder interferometer arrays,combining traditional artificial neural network algorithms and the optimized activation function.The main content of this paper includes the following parts:First,the concept,characteristics and significance of artificial neural networks are explained.By exemplifying the bottlenecks of the electronic integrated chips for AI chips,the concept of optical neural networks is introduced.Then,an optical neural network algorithm is constructed based on the artificial neural network algorithm,and the forward and back propagation in the algorithm are derived in detail.The influence of the different construction methods of feature input vectors and the selection of the activation function on the recognition accuracy are analyzed.Next,the photon integration technology and integrated optical waveguide theory are explained in detail,and the advantages and disadvantages of different materials constituting optical waveguides are analyzed.Derive the transmission matrices of the 3dB directional coupler and the Mach-Zehnder interferometer,determine the structure parameters of the Mach-Zehnder interferometer,and construct a silicon-based programmable nano-photon processor that can realize a unitary matrix function in detail.In Rsoft-BeamPROP,an optical neural network is constructed based on the Mach-Zehnder interferometer array.Then,the MNIST data set is used for global independent simulation,and the simulation results verify the feasibility of digital recognition by optical neural network.The scalability of the photon neural network is verified with the MNIST and FASHIONMNIST data sets.When the input port is increased,the recognition accuracy of both is improved.The influence of silicon waveguide loss on recognition accuracy is further discussed.Finally,the verification research in this paper is summarized,and the future prospects are combined with experimental device construction and photonic integrated chip structure.