Optical neural network chip construction method and system for reducing number of optical devices

The invention provides an optical neural network chip construction method and system for reducing the number of the optical devices, the number of redundant optical devices in an optical chip is approximately compressed through low rank by means of cooperation of a chip structure and a training method, and meanwhile, the recognition accuracy of a neural network is guaranteed based on a plurality of optimization means. The optical neural network chip construction method for reducing the number of optical devices comprises the steps of weight matrix preprocessing, unitary preserving training, device pruning and functional optical device network construction. Therefore, through the cooperation of the structure and the algorithm, under the condition that the recognition rate is basically kept unchanged, the use number of optical devices is greatly reduced. 本发明提出一种减少光器件数量的光神经网络芯片构建方法和系统，通过芯片结构与训练方法协同的方法，通过低秩近似压缩光芯片中冗余光器件的数量，同时基于若干的优化手段，保证神经网络的识别准确率。本发明所提出的减少光器件数量的光神经网络芯片构建方法包括权重矩阵预处理、酉保持训练、器件剪枝和功能光器件网络构建。因此，本发明通过结