Noisy-Label Learning

Noisy-label learning studies how models can learn when training labels are wrong, uncertain, or supplied by annotators with differing reliability. Common approaches include robust loss functions, small-loss sample selection and co-teaching, modeling annotator or label-noise processes, and detecting, reweighting, or correcting suspicious labels; newer work also adapts these ideas to language models, preference data, and federated learning. The field combines theoretical analyses of learnability under specific noise assumptions with empirical tests on benchmarks and naturally noisy datasets. A central tension is that filtering or correcting labels can reduce harmful supervision but also discard difficult, informative examples, and methods that work for one noise structure or data setting may fail in another.