Pseudo-Labeling

Pseudo-labeling studies how to use model-generated labels to learn from unlabeled examples, making it useful when human annotation is scarce or expensive. Common approaches include teacher–student self-training, confidence thresholding, consistency training, and iterative refinement, with applications spanning image segmentation, speech recognition, domain adaptation, and graph learning. Recent methods add calibration, uncertainty estimates, foundation-model cues, or temporal and multimodal agreement to decide which predictions to trust. The central tension is that more pseudo-labels increase coverage but can also amplify errors and confirmation bias, so gains depend on label quality, distribution shift, and the task.