Weakly Supervised Learning

Weakly supervised learning studies how to train useful models when labels are missing, noisy, ambiguous, or available only for groups of examples rather than individual instances. Common approaches include multiple-instance learning, positive-unlabeled and partial-label learning, learning from label proportions, and pseudo-labeling or combining heuristic labeling functions. These methods support tasks such as finding objects in images with only image-level labels, locating events in videos with sparse timestamps, and classifying medical scans from patient-level diagnoses. A central challenge is that weak labels can encode bias or shift across settings, so reliable evaluation and deciding when a label or localized prediction can be trusted remain open concerns.