Noisy Pseudo Label

Noisy pseudo-labeling in semi-supervised and unsupervised learning focuses on leveraging unlabeled data by assigning pseudo-labels, but addresses the challenge of inaccurate labels degrading model performance. Current research emphasizes refining pseudo-labels through techniques like dual prototypes, online mining, noise correction losses, and confidence-based filtering, often within teacher-student or co-training frameworks. These advancements improve the robustness of various machine learning models, particularly in scenarios with limited labeled data, impacting diverse applications such as object detection, semantic segmentation, and person re-identification. The ultimate goal is to effectively utilize unlabeled data while mitigating the negative effects of noisy pseudo-labels to achieve performance comparable to fully supervised methods.