Learning with Noisy Labels

Learning with noisy labels asks how models can learn useful decision rules when training targets disagree with the underlying truth, as often happens with crowd annotations, medical labels, and automatically generated supervision. Common approaches include robust losses and noise-transition models, sample selection and reweighting, co-teaching, label correction, and self-supervised or representation-based methods that identify suspicious examples without trusting their labels. Methods are tested on controlled synthetic corruption and real-world benchmarks such as CIFAR and Clothing1M, as well as in medical imaging, graphs, and other domains where noise may depend on the example or annotator. A central challenge is that noise structure and data difficulty matter: small-loss filtering can discard hard but correct examples, and success on closed-set synthetic benchmarks may not transfer to open-set noise or deployment settings without clean labels for evaluation.