Home | Publications | Wim25

Representing and Quantifying Predictive Uncertainty in Machine Learning

MCML Authors

Lisa Wimmer

Dr.

* Former Member

→ Group Bernd Bischl
Statistical Learning and Data Science

Abstract

This dissertation studies predictive uncertainty estimation in machine learning as a compositional problem of representation and quantification. It advances Bayesian methods for neural networks by exploiting posterior structure and symmetry to improve inference and uncertainty representation, and critically examines common uncertainty measures, proposing more principled alternatives. The work further demonstrates how reliable uncertainty estimates can improve downstream tasks such as adaptive data acquisition and self-supervised learning, highlighting both their practical value and remaining challenges. (Shortened).

phdthesis Wim25