Radial-Angular Geometry for Reliable Update Diagnosis in Noisy-Label Learning
The paper addresses a gap in noisy-label learning by shifting reliability estimation from forward signals like loss to analyzing parameter updates. It proposes Relative Geometric Conflict (RGC), which compares observed-label gradients against a reference gradient from an EMA teacher. This helps distinguish reliable hard-clean updates from conflicting updates caused by mislabeled data, leading to better sample preservation and accuracy.