Aravind Eye Hospital
Facing Failure to Generalize: Why some AI models exhibit underspecification.
The same models trained on the same data may show the same performance in the lab, and yet respond very differently to data they haven’t seen before. New work finds this inconsistency to be pervasive.