中国激光, 2017, 44 (4): 0406003, 网络出版: 2017-04-10
光子脉冲神经元权重器件的研制
Development of Weighting Device for Photon Spiking Neuron
光学器件 可调光衰减器 可塑性学习机制 光子脉冲神经元 神经网络 optical devices variable optical attenuator plasticity learning mechanism photon spiking neuron neural network
摘要
互连的光子脉冲神经元通过权重器件联系在一起, 为了实现神经网络的大规模计算, 权重器件的实现至关重要。利用微机电系统可调光衰减器(VOA), 研制了一种可以自动调节光子脉冲神经元的权重器件。该权重器件包括VOA、光电探测器、单片机、模数转换器、数模转换器和放大器等模块, 可以根据接收的光信号快速计算查表, 可对VOA的衰减值进行实时在线调整。该权重器件效率高, 且容易实现。该权重器件配合脉冲时间依赖的可塑性(STDP)光路使用, 可以实现光子脉冲神经元的STDP学习机制。当STDP曲线窗口高度为0.2时, 对权重器件进行了测量, 实现了4种STDP学习。实验测量结果与理论计算结果一致。
Abstract
Photon spiking neurons are connected by weighting devices, so the implementation of weighting devices is critical for realizing the large-scale computation of neural networks. Based on the variable optical attenuator (VOA), a weighting device for automatically adjusting photon spiking neuron is developed. The weighting device includes VOA, photoelectric detector, single chip, analog to digital converter, digital to analog converter, amplification module, etc. The weighting device can quickly calculate and look up table, and the attenuation values of VOA can be adjusted online based on the received optical signal. The weighting device has the advantages of high efficiency and easy implementation. When we combine the optical spike-timing-dependent plasticity (STDP) circuits and the proposed weighting device, STDP learning mechanisms for photon spiking neural can be achieved. The weighting device is detected when the height of STDP curve window is 0.2, and four STDP learning curves are obtained. The experimental results are consistent with the theoretical computation results.
宋晓佳, 王智, 李强, 孙翀翚, 乐燕思, 崔粲, 吴重庆, 刘彪. 光子脉冲神经元权重器件的研制[J]. 中国激光, 2017, 44(4): 0406003. Song Xiaojia, Wang Zhi, Li Qiang, Sun Chonghui, Le Yansi, Cui Can, Wu Chongqing, Lin Biao. Development of Weighting Device for Photon Spiking Neuron[J]. Chinese Journal of Lasers, 2017, 44(4): 0406003.