Reproduce “Loss Knows Best: Detecting Annotation Errors in Videos via Loss Trajectories” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Praditha Alwis, Soumyadeep Chandra, Deepak Ravikumar, Kaushik Roy
The authors introduce Cumulative Sample Loss (CSL), a method for flagging video labels for review using a model’s losses across training checkpoints. They train on reference videos, then score disjoint audit videos with frozen checkpoints. On synthetically disordered Cholec80 test videos, CSL reaches 78.5 AUC, versus 44.6 for final-epoch loss; this evidence is limited to synthetic corruptions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.