Holistic Reliability Propagation: Decoupling Annotation and Prediction for Robust Noisy-Label
For multimedia classification with noisy labels, HRP offers a more robust approach than methods that conflate annotation and prediction reliability.
HRP decouples annotation and prediction reliability for noisy-label learning, using meta-learned scalars to route them to separate objectives. It improves average accuracy over strong baselines and stays competitive at high noise rates.
Learning with noisy labels in multimedia classification often combines external annotations and model predictions into a single reliability weight, even though the two sources can fail for different reasons. We instead estimate disentangled reliabilities: bilevel meta-learning produces two batch-normalized scalars per sample, alpha for the given label and beta for the pseudo-label, without constraining them to sum to one. Holistic Reliability Propagation (HRP) then routes them to different objectives, using reliability-aware Mixup with global gating on the input branch and beta-gated pseudo-label positives on the contrastive branch. On synthetic and real-world benchmarks, HRP improves average accuracy over strong baselines and remains competitive at the highest noise rates.