光学神经网络智能处理:技术演变与未来展望  

Intelligent processing of optical neural networks: Technological evolution and future prospects

作　　者：李俊燊 孟祥彦 石暖暖 李伟[1,2,3] 祝宁华 李明[1,2,3] Junshen LI;Xiangyan MENG;Nuannuan SHI;Wei LI;Ninghua ZHU;Ming LI(Key Laboratory of Optoelectronic Materials and Devices,Institute of Semiconductors,Chinese Academy of Sciences,Beijing 100083,China;School of Materials Science and Optoelectronic Technology,University of Chinese Academy of Sciences,Beijing 100190,China;School of Electronics,Electrical and Communication Engineering,University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院半导体研究所光电子材料与器件重点实验室,北京100083 [2]中国科学院大学材料科学与光电技术学院,北京100190 [3]中国科学院大学电子电气与通信工程学院,北京100049

出　　处：《航空学报》2024年第20期61-83,共23页Acta Aeronautica et Astronautica Sinica

基　　金：中国科学院青年创新促进会(2022111);国家自然科学基金(62235011,62075212,61925505)。

摘　　要：卷积神经网络(CNN)以其卓越的特征提取能力,在人脸识别、图像分类、机器视觉、医学成像以及航空航天等领域展现出广泛的应用前景。然而,传统电学智能处理芯片受摩尔定律的制约,难以满足CNN算力需求的持续增长。光波以其超大宽带和超低损耗等特性,以光学或电学高维调控结构为基本单元,通过光的受控传播实现计算,是支撑下一代人工智能高算力需求的颠覆性技术。通过综述光学卷积神经网络的研究进展和技术突破,总结其发展的整体趋势,探讨未来需要解决的技术问题,并对光学卷积神经网络的应用前景进行展望。Convolutional Neural Network(CNN)has shown a wide range of applications in the fields of face recognition,image classification,machine vision,medical imaging,and aerospace due to its excellent feature extraction capabilities.However,traditional electrical intelligent processing chips are restricted by Moore’s law,and is difficult to meet the continuous growth of CNN computing power demand.With its characteristics of ultra-large broadband and ultra-low loss,light wave is a disruptive technology that supports the high computing power demand of the next generation of artificial intelligence.With optical or electrical high-dimensional control structure as the basic unit,it can realize computing through controlled propagation of light.In this paper,the research progress and technological breakthroughs of optical convolutional neural networks are reviewed.The overall trend of their development and the technical problems that need to be solved in the future are summarized.The prospects of optical convolutional neural networks for application are also discussed.

关 键 词：卷积 张量 神经网络 人工智能 深度学习 

分 类 号：V247[航空宇航科学与技术—飞行器设计] TN256[电子电信—物理电子学]