Leveraging Predictive Uncertainty to Enable Reliable Deployment of AI Models

This talk explores a unified theoretical framework for uncertainty quantification in machine learning, extending traditional methods to modern applications like generative modeling. By generalizing the bias-variance decomposition using proper scores, it introduces Bregman Information as a key component for epistemic uncertainty estimation. The session also covers uncertainty calibration, highlighting relationships between different calibration metrics. Finally, a novel approach for uncertainty-aware performance monitoring is presented, improving reliability under distribution shifts through active labeling interventions.

Organized by:

LMU Munich School of Management