中国激光, 2017, 44 (5): 0508002, 网络出版: 2017-05-03
反脉冲时间依赖可塑性学习机制的光学实现
Optical Implementation of Anti-Spike-Timing-Dependent Plasticity Learning Mechanism
非线性光学 神经拟态计算 反脉冲时间依赖可塑性 半导体光放大器 光子脉冲神经元 nonlinear optics neuromorphic computing anti-spike-timing-dependent plasticity semiconductor optical amplifier optical spiking neuron
摘要
突触可塑性为神经网络的学习机制提供了基础。基于单个半导体光放大器(SOA)的非线性偏振旋转(NPR)和交叉增益调制(XGM)效应实现了反脉冲时间依赖可塑性(anti-STDP)学习机制。通过调整SOA驱动电流,可以实现长时程增强窗口(LTP)和长时程抑制窗口(LTD)的高度和宽度调整,能更好地模拟神经网络。实验测量得到的anti-STDP曲线与生物系统中测量得到的学习曲线相吻合。使用该anti-STDP光路得到的学习曲线的时间窗口约为几百皮秒,其速度是人类大脑STDP学习机制的108倍。由于该anti-STDP光路系统简单,且SOA易于与其他器件集成,该anti-STDP光路可以用于实现大规模超快神经拟态计算系统。
Abstract
Synaptic plasticity provides the basis for learning mechanism in neural network. Anti-spike-timing-dependent plasticity (anti-STDP) learning mechanism is implemented by the nonlinear polarization rotation (NPR) and cross-gain modulation (XGM) based on single semiconductor optical amplifier (SOA). By adjusting the drive current of SOA, the weight and height of long-term potentiation (LTP) and long-term depression (LTD) windows can be adjusted to better mimic the neural network. The anti-STDP learning curves measured by the experiment closely resemble the learning curves measured by the biological system. Using the proposed anti-STDP optical circuit, the time window of anti-STDP learning curves is about several hundred picoseconds, which is 108 times faster than the speed of STDP learning mechanism in human brain. Since the proposed anti-STDP optical circuit is simple, and SOA can be integrated with some other devices easily, it can be used to realize large-scale and ultrafast neuromorphic computing systems.
李强, 王智, 崔粲, 乐燕思, 宋晓佳, 孙翀翚, 刘彪, 吴重庆. 反脉冲时间依赖可塑性学习机制的光学实现[J]. 中国激光, 2017, 44(5): 0508002. Li Qiang, Wang Zhi, Cui Can, Le Yansi, Song Xiaojia, Sun Chonghui, Liu Biao, Wu Chongqing. Optical Implementation of Anti-Spike-Timing-Dependent Plasticity Learning Mechanism[J]. Chinese Journal of Lasers, 2017, 44(5): 0508002.