光 计 算 和 光 电 智 能 计 算 研 究 进 展

Nan Zhang*, Zhiqi Huang*, Zian Zhang, Cong He, Chen Zhou, Lingling Huang*, Yongtian Wang

*此作品的通讯作者

科研成果: 期刊稿件文献综述同行评审

摘要

Significance Artificial intelligence (AI) is one of the most extensively investigated fields currently and has been widely applied in various fields such as agriculture, healthcare, home automation, and military. As its applications become increasingly complex, AI demands higher requirements on computing hardware in terms of computing power and energy efficiency. Currently, mainstream AI computations rely on integrated circuit chips such as central processing units (CPUs) or graphics processing units (GPUs), which are designed and manufactured using conventional microelectronic technologies, thus limiting their computing performance by the level of hardware-circuit integration density. According to Moore’s law, the integration density of circuits approximately doubles every 18 to 24 months. As Moore’s law reaches its limits and the demand for computing power continues to increase, researchers should develop transformative technologies that can overcome these challenges and shift computing to new paradigms. Optical (or photonic) computing, which uses optical fields as information carriers and optical devices to perform computations, has emerged as a promising and innovative field that can revolutionize various aspects of computing and information processing. It offers parallel and high-speed computation with significantly less energy consumption. Recently, interest toward optical computing and its interdisciplinary applications, i.e., platforms, architectures, integrable hardware and protocols for storage, encryption, and data and signal processing, has increased. Optical computing paradigms can be categorized into two main types: optical operators and optical neural networks. For optical operators, the optical element functions as an independent optical computing unit to perform a specific operation. It can be combined with back-end electronic calculation components to form an optoelectronic intelligent computing architecture that jointly performs complex functions. For optical neural networks, similar to artificial neural networks, optical components are used to construct optical neural networks, thereby directly achieving complex functions. This review comprehensively describes the operating principles, characteristics, and system-architecture features of the two major types mentioned above. Finally, it provides an outlook on the challenges and future development trends of optical computing and optoelectronic intelligent computing. Progress This article comprehensively reviews the research progress and challenges in optical computing. Based on the functionalities achievable by optical computing components, the novel hardware architectures can be classified into two categories: optical operators and optical neural networks. First, this review summarizes the methods for implementing optical operators with different functionalities, such as optical convolution, optical Fourier transform, and differentiator (Figs. 3 and 4). Subsequently, for optical neural networks, we systematically review optical implementation methods corresponding to the two core processes (i. e., optical linear operation and nonlinear activation function) of optical neural networks. For the optical linear operation, we introduce plane light conversion (PLC) (Fig. 6), Mach-Zehnder interferometer (MZI) (Fig. 7), and wavelength division multiplexing (WDM) methods (Fig. 8). Table 1 shows a comparison of different implementations for the typical optical linear operations. Additionally, a specific optical linear operation using phase change materials (PCMs) (Fig. 9) is discussed. For the nonlinear activation function, we discuss various methods associated with free-space optical neural networks (Fig. 10) and on-chip optical neural networks (Figs. 11 and 12). Table 2 lists the all-optical and electro-optic types that achieve nonlinearity, in addition to a comparison between them in terms of reconfigurability, power consumption, speed, and other aspects. In general, a well-performing network requires deliberate training. Herein, we analyze and discuss the training methods for optical neural networks, including offline training, online training, and the transition method (Figs. 14 and 15). Finally, we discuss the wide range of applications realizable by optical neural networks, including image processing, pattern recognition, optimization, and quantum computing (Fig. 16). Conclusions and Prospects Optical computing paradigms fully combine the advantages of multidimensional multiplexing, large bandwidths, and low power consumption of optics. Currently, the optoelectronic hybrid computing architecture is a universal morphology. Therefore, to further enhance the performance of optoelectronic hybrid computing architectures, comprehensive optimizations are required at both the hardware and algorithm levels. This involves integrating the flexibility of existing electronic computing with the bandwidth and speed of optical computing while minimizing the energy consumption associated with optoelectronic conversion, thereby preserving the low-power characteristics of optical approaches. In the future, integrated optical computing systems will become a development trend. Based on silicon photonics platforms, integrated computing systems provide numerous advantages such as compatibility with semiconductor processing and high integration density, thus rendering them the preferred solution and affording a diverse array of applications. Hence, heterogeneous integration processes must be improved to further increase system integration density and achieve device miniaturization. Additionally, all-optical architecture optical computing technologies should be investigated. This includes investigating nonlinear optical device schemes to realize all-optical neurons, analyzing optical interconnect technologies for flexible data transmission, and identifying optoelectronic memristor devices that can achieve low-power memory and real-time information processing. Whether in all-optical architectures or optoelectronic hybrid architectures, optical computing technologies should prioritize solving practical tasks. Therefore, reconfigurable optical devices must be analyzed to flexibly address various problems. Interdisciplinary collaborations and investments in cutting-edge technologies are expected to be key in revealing the full potential of optical computing and ushering in a new era of AI.

投稿的翻译标题Advances of Optical Computing and Optoelectronic Intelligent Computing
源语言繁体中文
文章编号1800001
期刊Zhongguo Jiguang/Chinese Journal of Lasers
51
18
DOI
出版状态已出版 - 9月 2024

关键词

  • artificial intelligence
  • optical computing
  • optical neural networks
  • optical operators
  • optoelectronic intelligent computing architecture

指纹

探究 '光 计 算 和 光 电 智 能 计 算 研 究 进 展' 的科研主题。它们共同构成独一无二的指纹。

引用此

Zhang, N., Huang, Z., Zhang, Z., He, C., Zhou, C., Huang, L., & Wang, Y. (2024). 光 计 算 和 光 电 智 能 计 算 研 究 进 展. Zhongguo Jiguang/Chinese Journal of Lasers, 51(18), 文章 1800001. https://doi.org/10.3788/CJL240780