Julia Machnio

A Mechanism-Driven Theory of Phase Transitions in Active Learning

Julia Machnio, Mads Nielsen, Mostafa Mehdipour Ghazi

Active learning (AL) performance is known to be budget-dependent, yet regimes are typically defined by heuristic label counts that fail to generalize across datasets or architectures. We characterize AL dynamics by reframing budget regimes as shifts in the dominant generalization…Read more