Unfolding Sample-Level Difficulty for Adaptive Pseudo-Labeling in Semi-Supervised Learning

Semi-supervised learning (SSL), which utilizes a large number of unlabeled data incorporating few annotations, has gained considerable attention in the big data era. Within the domain, one of the most pivotal tasks lies in how to efficiently utilize unlabeled data. The prevalent approaches mainly rely on pseudo-labeling, where the model first predicts on unlabeled data, and then selects high-confidence predictions (i.e., pseudo labels) to be used in the next training iterations. Previous studies in pseudo-labeling methods have typically relied on fixed confidence thresholds for prediction outcomes or dynamically adjusted class confidence thresholds based on class-level performance. However, these methods largely ignore the inherent difficulty of the unlabeled data samples and their impact on pseudo-label confidence. To this end, we conduct a thorough investigation of mechanisms that adjust pseudo-label confidence according to sample difficulty and propose a novel SSL framework, called SampMatch. This framework facilitates the adaptive learning of pseudo-label confidence thresholds by unfolding the sample-level difficulty of each class. Additionally, SampMatch integrates a class balancing penalty mechanism based on the number of pseudo-labeled samples to further mitigate the potential for class prediction bias during model training. Extensive experiments on two public benchmarks demonstrate the effectiveness of our proposed SampMatch.