Improving Pseudo-Labeling and Representation Balance in Realistic Long-Tailed Semi-Supervised Learning

Résumé fourni par la source

Despite the remarkable progress of semi-supervised learning (SSL), its effectiveness under realistic long-tailed settings remains limited. In such settings, labeled data is severely imbalanced, while the distribution of unlabeled data is unknown and often mismatched. Under these conditions, class imbalance inherently leads to biased decision boundaries during training, and distribution mismatch causes unreliable pseudo-labels that further exacerbate this bias. Moreover, realistic long-tailed semi-supervised learning suffers from representation imbalance in feature learning, where dominant classes occupy large regions of the feature space while minority classes become overly compact. To address these challenges, we propose PRB-SSL, a method for improving pseudo-label reliability and representation balance in realistic long-tailed semi-supervised learning. PRB-SSL is built upon a dual-branch framework. Specifically, a Biased Predictor adapts to the unlabeled data distribution to generate more reliable pseudo-labels under distribution mismatch, while a Balanced Predictor with decision-level rebalancing mitigates class-imbalance-induced boundary bias and enables balanced inference. Furthermore, PRB-SSL introduces learning status, a dynamic class-level measure, to regulate feature diffusion during semi-supervised learning, suppressing excessive expansion of well-learned classes while preserving exploration for under-learned ones, thereby alleviating representation imbalance. Extensive experiments on CIFAR-10-LT, CIFAR-100-LT, and STL-10-LT demonstrate that PRB-SSL consistently outperforms state-of-the-art methods under realistic long-tailed semi-supervised learning settings.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.