Loss Knows Best: Detecting Annotation Errors in Videos via Loss Trajectories

Name: Loss Knows Best: Detecting Annotation Errors in Videos via Loss Trajectories
Uploaded: 2026-09-06T07:06:55.210601+00:00
Duration: 501 s
Description: The episode discusses 'Loss Knows Best: Detecting Annotation Errors in Videos via Loss Trajectories,' a method for auditing video training data. The technique uses Cumulative Sample Loss (CSL) to diagnose annotation errors by tracking how consistently a model struggles with specific frames over time, proving data quality rather than assuming it.

The episode discusses 'Loss Knows Best: Detecting Annotation Errors in Videos via Loss Trajectories,' a method for auditing video training data. The technique uses Cumulative Sample Loss (CSL) to diagnose annotation errors by tracking how consistently a model struggles with specific frames over time, proving data quality rather than assuming it.

Open the full episode