Affinity-aware uncertainty quantification for learning with noisy labels

doi:10.1016/j.patcog.2025.112495

Affinity-aware uncertainty quantification for learning with noisy labels

2025-09-26

0

PRE

AI

DOI：10.1016/j.patcog.2025.112495

Original

Original request for help

Save

Abstract

En 中文

• We propose the affinity-based uncertainty quantification framework (AUQ) to address salient bias, where the uncertainty distribution is learned using dynamic prototypes to emphasize learning from hard samples. • We develop adaptive pseudo-label refinement and masking strategy to enhance pseudo-label quality and robustness. • We design an uncertainty loss based on Monte Carlo algorithm to improve model’s noise robustness.

Affinity-aware uncertainty quantification for learning with noisy labels

Abstract

Journal

Organization