#machine learning
Feb 2026
Loss Knows Best: Detecting Annotation Errors in Videos via Loss Trajectories
Cumulative Sample Loss (CSL) is computed as the mean annotation-conditioned loss of an audit frame across checkpoints trained on a *disjoint* reference set and acts as a dynamic fingerprint and captures the persistent disagreement between its annotation and learned visual-temporal structure.
Praditha Alwis, Soumyadeep Chandra, Deepak Ravikumar et al.
· arXiv.org · 0 citations