Title: Leveraging predictive uncertainty to enable reliable deployment of AI models
About the Talk:
This talk presents a comprehensive theoretical framework for uncertainty quantification in machine learning based on proper scores and its application across various real-world tasks. It bridges traditional uncertainty quantification methods from classification and regression to more modern applications such as generative modeling. To this end, the work extends the traditional bias-variance decomposition beyond mean squared error to strictly proper scores, demonstrating how Bregman Information naturally emerges as the variance term. This provides a principled approach to epistemic uncertainty quantification across different tasks.
Additionally, the talk introduces a unified approach to uncertainty calibration through proper scores, illustrating how different calibration metrics relate to one another. It explores how proper calibration errors, directly derived from proper scores, can be reliably estimated and leveraged in various downstream tasks.
Finally, a novel approach for uncertainty-aware performance monitoring is presented, enabling reliable performance estimates under various distribution shift scenarios by guiding active labeling interventions.
About the Speaker: Florian Buettner is a professor at Goethe-University Frankfurt and the German Cancer Consortium (DKTK)/German Cancer Research Center (DKFZ). Having earned his PhD in physics he spent several years as a postdoc and principal investigator at the Helmholtz Center for environmental health Munich and the European Bioinformatics Institute, Cambridge. Prior to joining Frankfurt University, he worked as a guest scientist at the Helmholtz Center Munich, in addition to his appointment as senior scientist for industrial AI and probabilistic machine learning at Siemens AG. His research interests are focused on the intersection of multi-omics bioinformatics, machine learning and oncology.