Curtis Northcutt

First author of pervasive-label-errors (NeurIPS 2021, with Anish Athalye and Jonas Mueller), the paper that measured how wrong the field’s standard test-set labels are and showed that correcting them can reverse a model ranking.

The method the paper rests on is confident learning — estimating which labels are wrong from a model’s own predicted probabilities, then putting the flagged cases to human review. Nothing further is evidenced by the source held here.