| In recent years,artificial neural network ANN has gradually become a research hotspot and mainstream development direction in the field of artificial intelligence.With the development of optical devices,optical matrix operations,spatial optical interconnection,optical nonlinear operations and other technologies gradually become the stateof-art research,optical neural networks become a future alternative to neural networks in electronics and has great potential for computing solutions.Using diffraction,scattering,optical interference,optical Fourier transform and optical wavelength division multiplexing,many scholars at home and abroad have recently proposed a variety of architectures for optical neural networks,such as fully diffraction optical neural networks,optoelectronic hybrid neural networks,neuromimetic optical on-chip networks,etc.,which have greatly contributed to the development of optoelectronic intelligent computing.In this paper,we make an in-depth analysis of diffraction optical neural network architecture among various optical neural networks,design a neural network optical computing model composed of multiple diffraction surfaces to replace the fully connected layer computation in traditional digital neural networks from diffraction theory,and use spatial light modulator(SLM)as the key device for diffraction neural network layer construction,design and build a diffraction neural network optical path system,use the network model and system for image classification recognition and verify the feasibility of using diffraction light propagation to complete the fully connected computation of the neural network layer under this scheme through experiments.The details are as follows:(1)Diffraction propagation instead of digital neural networks for fully connected layer computational analysis.From the Fresnel-Kirchhoff diffraction theory,the physical properties of diffraction propagation in space under near-field as well as far-field conditions are analyzed,and the behavior of diffraction unit propagation in diffraction nerves is determined to be analyzed on the basis of the Rayleigh-Sommerfeld diffraction formula.By comparing the similarities between the fully connected model of electronics and diffraction propagation,the diffraction formula is discretized and analyzed,and a mathematical model is used to describe the propagation of individual diffraction unit in the diffraction plane,providing a mathematical guide for the implementation of diffraction neural network modeling.(2)Diffraction field analysis and diffraction neural network modeling of SLM devices.A novel optical device,liquid crystal spatial light modulator(LC-SLM),for spatial light field control is discussed.The physical characteristics and operating principle of the device are analyzed,and the device amplitude and phase control capabilities are quantified and calibrated.At the same time,the diffraction field modeling analysis is performed according to the SLM device parameters to determine the mathematical modeling method of diffraction surface propagation,and finally the multilayer pure phase diffraction light neural network model is designed according to the principle of diffraction calculation and performs the image classification and recognition task,and the final classification rate of MNIST dataset reaches 87.6%,which proves the computational feasibility of diffraction light propagation as a neural network.(3)To improve the performance of the network model further,a nonlinear photoelectric activation function PRe LU function is designed for the diffraction optical neural network by using CMOS for threshold truncation of the detected optical field and linear response capability of light intensity.By improving the network design,the diffraction model achieves 94.1% correct rate for MNIST image dataset classification application and 92.1% correct rate for complex image classification(Fashion-MNIST dataset),with significant performance.(4)Optical implementation of diffraction optical neural network.According to the improved nonlinear diffraction neural network model and training parameters,the SLM diffraction device is used as the core to design and build the corresponding optical path system.After completing the simulation analysis and assembly adjustment of the optical path system,the optical classification prediction experiments for the above two image datasets are carried out.Finally,MNIST and Fashion-MNIST datasets have 91%(500 images)and 81.7%(1000 images)accuracy in optical classification.The proposed design scheme has good image classification performance compared with similar works at home and abroad,the simulation and optical experiments demonstrate the feasibility and effectiveness of the scheme for multilayer fully connected neural network optical computing. |
