Physical Aware Clustering Training Method for Integrated Photonic Convolution Neural Network with Nonlinear Distributed Weights
Yue Jiang, Wenjia Zhang, Xuying Liu, Wenyu Zhu, Zuyuan He · 2022
The computing power of the Photonic Convolution Neural Network (CNN) have achieved Tera-level operations per second (OPS) for supporting machine vision algorithms. However, restrained by the nonlinear characteristics of weighting devices, it is hard to realize convergent training of Photonic CNN under poor controlling precision. In this paper we proposed a Physical Aware Clustering training method where the Physical Aware Cluster Quantizer is embed with the straight through estimator (STE) algorithm for integrated Photonic CNN with quantized and nonlinear distributed weights. Experiment shows that, by employing the upgraded STE methods (namely, STE plus), the Photonic CNN using micro ring weighting bank and PAM4 controlling modules achieves the nearly 99.3% accuracy ratio for Fashion MNIST recognition task, whereas only 41.5% of that by using baseline STE training algorithms with the same physical devices.