光神经拟态系统能以比生物大脑快几百万至十亿倍的运行速度模拟神经拟态算法, 优于电神经拟态硬件系统, 且其可胜任比传统光计算更复杂的计算任务。光神经拟态计算探索超快光脉冲信号的自适应性、稳健性和快速性, 能够避免传统数字光计算的芯片集成及模拟光计算的噪声积累等问题。本文报道了光子神经拟态信息处理的发展历程, 并从光子神经元, 光脉冲学习算法以及可集成光学神经拟态网络框架等方面介绍了光神经拟态计算的关键理论和技术。阐述了光神经拟态计算研究的必要性及存在的问题, 展望了其潜在的应用前景。

Photonic neuromorphic system can simulate the neuromorphic algorithms with a speed of millions or billions multiples faster than the biological counterpart, which is better than electronics neuromorphic hardware system. It can be also capable of processing computing tasks more sophisticated than traditional optical computations. The photonic neuromorphic computing, exploring the adaptability, robustness and rapidity of ultrafast optical pulses, can overcome the scaling of digital optical computation and the noise accumulation of analog optical computation. The research progress on photonic neuromorphic computing is reported. Some essential theories and technologies, including photonic neuron, learning algorithms based on optical spiking pulses and the integrated photonic neuromorphic network framework, are introduced respectively. The necessity of research on photonic neuromorphic computing and its problems are discussed to present its potential applications in future.