Certifying Whether to Adapt: Baseline-Relative PAC-Bayes-MMD Guarantees for Test-Time Adaptation
The paper develops a PAC-Bayesian framework to decide whether to adapt a model during test-time adaptation. It uses Maximum Mean Discrepancy (MMD) between source and target distributions to quantify uncertainty, interpreting MMD balls as imprecise probability sets. This provides formal guarantees on the target risk, showing how MMD shift affects generalization bounds.