中心博士生徐绍夫的工作——针对光子卷积神经网络的光子分块方法的相关成果近期发表在Optics Letters期刊上，该工作得到了国家重点研发计划（2019YFB2203700）、国家自然科学基金（61822508）的部分资助。推广化的矩阵乘法（Generalized matrix multiplication）是计算卷积神经网络的一项重要方法。为了实现推广化的矩阵乘法，必须对输入数据完成分块操作。徐绍夫博士生通过将分块操作的数学原理进行一维序列化变换，推导出利用光延时线进行光子分块操作的方法，即“光子分块方法（Optical patching scheme）”。并且利用波分复用技术将多个波长上加载的数据通过一个光延时线组合同时完成分块操作。通过分立器件的实验验证了光子分块方法的可行性。上述的光子分块方法在不降低系统计算能力的前提上大幅减少光子卷积神经网络中所使用的调制器的数量，有助于提升光子神经网络系统的能量效率。此外，相对于以往的光子卷积神经网络工作，本文提出的光子分块技术使用的延时线总长度降低了数十倍。
摘要: Recent progress on optical neural networks heralds a new future for efficient deep learning accelerators, and novel architectures of optical convolutional neural networks provide potential solutions to the widely adopted convolutional models. In so-far optical convolutional neural networks, the data patching (a necessary process in the convolutional layer) is mostly executed with electronics, resulting in a demand for large input modulator arrays. Here, we experimentally demonstrate an optical patching scheme to release the burden of electronic data processing and to cut down the scale of input modulator array for optical convolutional neural networks. Optical delay lines replace electronics to execute data processing, which can reduce the scale of input modulator array. The adoption of wavelength-division multiplexing enables a single group of optical delay lines to simultaneously process multiple input data, reducing the system complexity. The optical patching scheme provides a new solution to the problem of data input, which is challenging and concerned in the field of optical neural networks.