Optical Prediction Based on Less Selectively-Labeled Samples and Cross-Semi-Supervised Learning

Abstract

A large amount of labeled data is usually utilized to fully training forward prediction models, which results in a heavy labeling burden. Here, under constrained annotation resources, a scheme combining less selectively-labeled samples with cross-semi-supervised learning is proposed to accurately predict directional scattering from nanostructures. It is found that when only one-third dataset are labeled by numerical simulation, the prediction accuracy in this scheme is comparable to that of the conventional method based on fully-labeled data. Our findings greatly reduce the labeling cost for deep learning tasks in the field of nanophotonics and provide a new way to efficiently utilize limited data resources.