1 paper across 1 session
We analyze imbalanced training loss, showing that gradient descent dynamics can gradually reduce bias and recover minority-specific features with longer training.