MMD-Balls as Credal Sets: A PAC-Bayesian Framework for Epistemic Uncertainty in Test-Time Adaptation
Researchers have developed a PAC-Bayesian framework to quantify epistemic uncertainty in test-time adaptation (TTA) methods. This framework uses maximum mean discrepancy (MMD) between source and target distributions to derive generalization bounds. By interpreting MMD-balls as credal sets, the approach separates epistemic from aleatoric uncertainty, offering a principled way to decide when adaptation is beneficial. AI
IMPACT Provides a theoretical foundation for understanding and quantifying uncertainty in models adapting to new data distributions.