Edge of Stability selectively shapes learning across data distribution

Arxiv CS.LG (machine learning)

The paper demonstrates that the edge of stability (EoS) effect is selective, not global, redistributing learning across training data subsets. This selective dynamics amplifies progress on certain examples while slowing learning on others, challenging existing theories.

Arxiv CS.LG (machine learning)·Shauna KwagAnakha GaneshTomaso PoggioPierfrancesco Beneventano

7 days ago