Pseudo-labeling is an attractive approach to semi-supervised image classification because it turns unlabeled images into ordinary supervised training examples. Its simplicity is also its main weakness: confident early mistakes can be reinforced by later optimization, especially when the labeled split is small, classes are imbalanced, or the unlabeled pool contains shifted examples. This paper presents Gated Pseudo-Labeling (GPL), a conservative training procedure that admits an unlabeled example only when an exponential-moving-average teacher is confident, two weakly perturbed views agree on the class, and the predicted distributions remain close under perturbation. Accepted examples train the student on strongly augmented views, deferred examples remain eligible in later epochs, and a class-level acceptance ledger adjusts thresholds when some classes are persistently under-admitted. A controlled study on CIFAR-10, SVHN, STL-10, and a shifted unlabeled-pool variant shows that GPL improves low-label robustness relative to plain pseudo-labeling, temporal ensembling, and Mean Teacher baselines, while adding only a small overhead beyond a second weak-view teacher pass. The results suggest that pseudo-labeling should be treated as an admission-control problem, not only as a confidence-thresholding heuristic.
Saha et al. (Sat,) studied this question.