Research On Machine Translation Method Based On Optical Computing

With the continuous improvement of the complexity of artificial intelligence tasks and the continuous increase of big data business volume,the amount of interconnection and calculation of various super large and high concurrency data centers are exploding,and the amount of high-dimensional matrix calculation involved is increasing day by day,which promotes the higher and higher integration of computing units.However,the traditional electronic computing unit is limited by Moore’s law and can not further improve its performance.It is increasingly inadequate in terms of energy consumption,computing speed and concurrency,which brings severe challenges to the efficiency and stability of computing nodes.With the advantages of high concurrency,high speed,low energy consumption and anti electromagnetic interference,optical computing can perform high-dimensional matrix computing tasks in parallel at the speed of light,which can break through the bottleneck encountered by traditional electronic computing.It has broad research prospects in the interconnection of super large data centers,artificial neural networks and next-generation chip assembly.Aiming at the opportunities and challenges brought by optical computing in machine translation task,which involves a large number of high-dimensional matrix operations,this paper designs a machine translation system based on optical computing method and carries out experimental verification.The main research contents are as follows:Firstly,taking advantage of the direct two-dimensional image processing of optical computing,a multilingual text image translation system is designed to realize the translation from image to image.Based on the functional requirements of the system,it is divided into image feature recognition,feature compression coding,word vector translation and word vector correlation optimization,and the functions of the above modules are realized by different optical computing methods.Compared with traditional machine translation methods,the translation method based on optical computing can save at least 50%of the running time.With the increase of the dimension of the calculation matrix,the proportion of the saved running time and storage space will further increase.Secondly,build an optical experimental platform to realize machine translation,verify the feasibility of the simulation results and evaluate the computational performance of the experimental platform.The optimization scheme is designed for various errors caused by the limitation of experimental conditions.In addition,in order to make up for the additional time complexity brought by the error optimization scheme,a method of accelerating matrix loading on the hardware platform is adopted.The calculation accuracy of the experimental platform after error optimization is improved.Finally,according to the experimental results,we analyze the feasibility and performance differences of different optical computing methods in machine translation scenarios,and look forward to the research direction of optical computing in solving more machine translation tasks and the broad prospect of realizing all-optical machine translation system.