Pseudo Label Selection

Pseudo-label selection (PLS) in semi-supervised learning aims to improve model performance by carefully choosing which unlabeled data points to assign pseudo-labels to, thereby guiding model training with high-quality, informative examples. Current research focuses on developing sophisticated selection criteria that go beyond simple confidence thresholds, incorporating techniques like joint assessment of classification confidence and localization reliability, uncertainty-guided selection, and even learning-order-based approaches. Effective PLS is crucial for various applications, including temporal action localization, continual test-time adaptation, and semi-supervised domain generalization, leading to more data-efficient and robust machine learning models.