Label Distribution Learning

Label distribution learning studies how to train models from label probabilities or proportions, capturing ambiguity, annotator disagreement, and ordinal uncertainty that a single label hides. Common approaches use soft-target cross-entropy or KL divergence, probabilistic models of annotator counts, and distribution-aware curricula or distillation; applications include NLI, emotion recognition, medical imaging, and ordinal prediction. These methods are evaluated not only by classification scores but also by how well they match human distributions, uncertainty, and calibration. A central tension is that useful distributions require enough reliable annotations: soft targets often help when disagreement matters, while sampled hard labels can be competitive when votes are sparse.