Pseudo-label Selection for Deep Semi-supervised Learning
Shuangshuang Li, Zhihui Wei, Jun Zhang, Liang Xiao
Abstract
Deep semi-supervised learning is a hot research topic in recent years. The main challenges are the small sample learning and make full use of unlabeled data. Pseudo-labeling is a simple yet effective approach in semi-supervised learning. However, how to obtain high quality pseudo-labeled data is key issue. When pseudo-labeled data with low confidence level are used to train deep learning model will enlarge or spread classification errors. To overcome this issue, we propose a confidence-based pseudo-labeling method to choose high quality pseudo-labels with a novel uncertainty model. Specifically, we treat the uncertainty of pseudo-labels as a random variable and model proposed to estimate the confidence level of the pseudo-labels. The selected pseudo-labeled data are produced by minimizing the average risk criterion. Experimental results on the MNIST dataset show that the proposed method is able to improve the performance for deep semi-supervised learning.