On Pseudo-Labeling for Class-Mismatch Semi-Supervised Learning

Lu Han, Han-Jia Ye, De-Chuan Zhan

Abstract: Semi-Supervised Learning (SSL) methods have shown superior performance when unlabeled data are drawn from the same distribution with labeled data. Among them, Pseudo-Labeling (PL) is a simple and widely used method that creates pseudo-labels for unlabeled data according to predictions of the training model itself. However, when there are unlabeled Out-Of-Distribution (OOD) data from other classes, these methods suffer from severe performance degradation and even get worse than merely training on labeled data. In this paper, we empirically analyze PL in class-mismatched SSL. We aim to answer the following questions: (1) How do OOD data influence PL? (2) What are the better pseudo-labels for OOD data? First, we show that the major problem of PL is imbalanced pseudo-labels on OOD data. Second, we find that when labeled as their ground truths, OOD data are beneficial to classification performance on In-Distribution (ID) data. Based on the findings, we propose our model which consists of two components -- Re-balanced Pseudo-Labeling (RPL) and Semantic Exploration Clustering (SEC). RPL re-balances pseudo-labels on ID classes to filter out OOD data while also addressing the imbalance problem. SEC uses balanced clustering on OOD data to create pseudo-labels on extra classes, simulating the process of training with their ground truths. Experiments show that our method achieves steady improvement over supervised baseline and state-of-the-art performance under all class mismatch ratios on different benchmarks.