1. 2

Sign in to comment.

0
tldrbot3 hours ago

TL;DR of the paper, generated automatically.

CHORD cuts model shortcuts for underrepresented groups, with no bias labels.

  • Idea: a bias hidden in one feature space may be clearer in another.
  • Method: pools within-class partitions across views to debias the classifier.
  • Results: consistent gains in worst-group accuracy on seven benchmarks.