摘要
人工神经网络(ANN)学科的发展推动了信息处理技术的进步,目前,被称为第3代ANN的脉冲神经网络(SNN)以其更具生物可解释性,更适合ANN硬件实现的优势受到业界的广泛关注,并已成功应用于模式识别、医学成像和智能控制等多个领域。受制于“后摩尔时代”电子芯片的制程不断接近极限以及冯·诺依曼体系“存算分离”带来的性能瓶颈,低时延、低能耗、高带宽和高并行性的光子计算方案应用于SNN的硬件实现成为信息处理领域多学科融合的热门课题。文章介绍了光子SNN的起源,利用光学器件的特性实现神经元的行为和突触连接强度的变化进而实现SNN的研究历程和多种实现方案,总结了光子SNN目前存在的瓶颈和挑战,展望了光子SNN的未来发展趋势。
The development of Artificial Neural Network(ANN) discipline promotes the progress of information processing technology. At present, the Spiking Neural Network(SNN) known as the third generation ANN is widely attracted more attention in the industry because of its advantages of behaving more biologically interpretable and more suitable for the implementation of ANN hardware, and it has been successfully applied to pattern recognition, medical imaging, intelligent control and other fields. Constrained by the fact that electronic chip manufacturing process is constantly approaching the limit in the "Post-Moore Era" and the performance bottleneck brought by the "separation of storage and computing" in the von Neumann system, photonic computing solutions with low latency, low energy consumption, high bandwidth, and high parallelism applicated to the hardware implementation of SNN has become a hot topic of multi-disciplinary integration in the field of information processing. This paper introduces the origin of the photonic SNN, the research process and various implementation schemes that use the characteristics of optical devices to realize the behavior of spiking neuron and synaptic connection strength thus realizing the SNN, and summarizes the current bottlenecks and challenges of the photonic SNN. And the future development trend of photonic SNN is also prospected.
作者
柯特
朱盈
彭楚宇
胡晓
肖希
KE Te;ZHU Ying;PENG Chu-yu;HU Xiao;XIAO Xi(State Key Laboratory of Optical Communication Technologies and Networks,China Information Communication Technologies Group Corporation,Wuhan 430074,China)
出处
《光通信研究》
2023年第1期17-31,共15页
Study on Optical Communications
基金
国家重点研发计划资助项目(2022YFB1806401)。
