中心博士生张俊峰的工作——A Large-Scale Photonic CNN Based on Spike Coding and Temporal Integration（基于脉冲编码和时间积分的大规模光子卷积神经网络）的相关成果近期被IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS期刊接收发表，该工作得到了国家重点研发计划（2019YFB2203700）、国家自然科学基金（T2225023）的部分资助。随着实时海量数据智能信息处理任务的不断增加，对新型智能优化硬件的需求也日益凸显。因此，我们提出了一种可重构的基础处理单元，用于构建大规模的光学卷积神经网络。该处理单元能够实时调整其突触权重，以适应不同的图像处理和分类任务。本研究采用了生物学中的“首次脉冲编码”方法，将脉冲编码应用于DFB-LD神经元，用于执行图像处理任务中的时间卷积运算。实验结果表明，我们提出的方案成功实现了大尺寸的卷积核操作，用于提取图像特征的时间卷积运算。此外，本研究还探究了DFB-LD神经元的时间脉冲成形技术，并建立了一种密集连接的全连接层。神经元突触权重能够以5 GHz的速度快速调整，实现了高精度的MNIST和Fashion-MNIST基准图像分类任务。因此，本研究提供了探索高性能、大规模、可重构的光学神经网络的新视角，并突显了类似神经元的光学模拟计算平台在实时和更复杂的智能处理网络方面的潜力。这些结果有望应用于未来智能处理器架构，实现实时和超高带宽数据的处理。
摘要: The real-time massive-data intelligent information processing tasks highlight a vital requirement for the novel, intelligent optimization hardware. Convolutional neural networks are highly capable of extracting the hierarchical feature map and enhancing recognition accuracy, with photonics-enabled implementations drawing considerable interest. Here, we propose a large-scale and reconfigurable photonic convolutional neural network (PCNN), based on a hardware-friendly distributed feedback laser diode (DFB-LD). Our approach applies biological time-tofirst-spike coding to a DFB-LD neuron to perform a temporal convolutional operation (TCO) for image processing tasks. In PCNN, experimental results demonstrate that we successfully implement the TCO to extract the image features with convolutional kernels of size 11 × 11. Furthermore, we investigate the temporal pulse shaping of a DFB-LD neuron to build a densely-connected fully connected layer, which synaptic weights can be rapidly adjusted at a rate of 5 GHz, achieving full MNIST and Fashion-MNIST benchmark image classification tasks, with classification accuracies of 98.56% and 87.48%, respectively. This work highlights the potential of neuron-like optical analog computing platforms for real-time and more complex intelligent processing networks.