Research Of Diffraction Optical Neural Network System Based On Optoelectronic Devices

In recent years,there has been an increase in data-driven applications in industrial,medical,civil,and military fields.The massive data computing makes it difficult for traditional silicon chip computing devices to meet the demand,prompting mankind to seek new solutions to meet the demand for high-speed interconnection and computing of massive data.Optical computing has attracted much attention due to its ultralow computing latency and extremely low energy consumption.Since the 21 st century,with the development of optoelectronic devices,various optical modulation devices and photodetectors have emerged,providing the possibility for the implementation and integration of optical computing.This paper follows the optical computation of fully connected neural networks.Fully connected neural networks are the beginning of modern deep learning models,and they are the most basic neural network models.The neuron outputs of the current layer are transmitted to all neurons of the next layer,similar to the Huygens-Fresnel principle,which states that each point of the wave front emits sub-waves that interfere with each other and superimpose to form a new wave front.In this paper,analyze the similarity between the propagation of light waves between multilayer media and the forward propagation process of multilayer fully connected neural networks according to the Huygens-Fresnel principle,and construct a reconfigurable adaptive multilayer diffractive optical neural network system,which can complete the computation of multilayer digital fully connected neural networks with high parallelism and low energy consumption of optical computation.Further,this paper verifies the application of the proposed reconfigurable adaptive multilayer diffractive optical neural network system for image classification.The main work and research results of this paper are as follows.(1)A comprehensive review of neural network computation in optical computing research and progress is presented.The existing optical computing schemes are organized and summarized,from basic neuron structures and classical fully-connected networks to convolutional neural networks(CNNs)and recurrent neural networks(RNNs).Focusing on the analysis and study of the optical computing implementation of the basic neuron structures of neural networks,the theoretical basis,technical route,and implementation methods of optical computing are introduced.This mainly summarizes the challenges of existing optical computing research used to realize neural network computation and proposes the improvement direction of the optical computing architecture used to realize neural network computation according to the new progress of optoelectronic device technology.(2)The reconfigurable optical neural network structure is designed based on the independent propagation characteristics of light for the current diffractive optical neural network architecture,which is difficult to flexibly adjust the number of network layers,the number of neuron nodes within the network layers,and the size of input and output data.By dividing the large-sized spatial light modulator and photodetector devices into multiple partitions for independent control,the optical computation of multilayer neural networks is realized,which can execute the computation of multiple neural network layers in parallel or multiple network layers of a single neural network in serial.In addition,the problem of inter-partition interference that may be caused by dividing large-sized spatial light modulators and photodetectors into multiple partitions for independent use is also solved.Based on the Huygens-Fresnel principle simulation,the interference problem between multiple partitions in the same optical plane is analyzed and a method to solve the inter-partition interference is given.(3)The effects of diffraction optical neural network interlayer distance,network layer size,and neuron node size on neural network performance are analyzed.A model of interlayer neuron data exchange in diffractive optical neural networks is constructed based on the Huygens-Fresnel principle.The relationship between the information transfer of neurons in diffractive optical neural networks and the interlayer distance of network layers,the size of network layers,and the size of neuron nodes is analyzed,and the optical structure of the computational optical path of reconfigurable optical neural networks and the division rules of multiple partitions are designed accordingly.The problem of flexibly adjusting the number of neuron nodes in network layers and the size of input and output data of existing multilayer neural networks is overcome,thus expanding the practicality of optical computation of neural networks.(4)A reconfigurable diffractive optical neural network computing unit has been constructed based on existing silicon-based liquid crystal spatial light modulation devices and photodetector devices.This constructed optical neural network computational unit can realize the computation of multilayer fully connected neural networks without the need to readjust the optical path,and it can be extended to different neural network structures and more network layers.A method for aligning the relative positions of the mainly optics using holograms has also been provided.Finally,on the MNIST dataset and Fashion-MNIST dataset,the diffractive optical neural network computational unit proposed in this paper running a four-layer fully connected network achieved 89% and81% classification accuracies,which are the same performance as various state-of-theart international optical neural networks currently available.