Semi-Supervised Learning
Also known as SSL
Overview
Semi-supervised learning studies how to combine a small labeled dataset with a larger unlabeled one, reducing annotation costs in tasks such as image segmentation, speech recognition, and graph classification. Common approaches include pseudo-labeling, teacher–student consistency training, graph-based label propagation, and generative or statistical methods that use unlabeled data to learn structure or improve inference. Current work also adapts these techniques to settings such as open-set recognition, class imbalance, domain shift, and foundation-model tuning. A central challenge is deciding which unlabeled examples and predictions are trustworthy: distribution mismatch, noisy pseudo-labels, and confirmation bias can make additional data harmful, so gains remain dependent on the task and data quality.
Latest papers
167 papers5 papers in the last four weeks, down 67% on the four weeks before. 0.0% of all new papers.