Evaluating Uncertainty and Robustness in Vision-Based Models

Traditional performance metrics such as accuracy provide only a limited view of machine learning models. Two models can achieve identical accuracy while differing substantially in reliability, for example under distribution shifts or when producing overconfident errors. As machine learning models are increasingly deployed in safety-critical domains like medical imaging, evaluating robustness and uncertainty becomes essential.

This thesis introduces two evaluation frameworks that enable rigorous analyses of model behavior. The first focuses on uncertainty estimation in semantic segmentation, systematically analyzing uncertainty types, components of uncertainty methods, and application to downstream tasks. The second focuses on robustness evaluation for Vision Language Models in medical visual question answering, incorporating realistic distribution shifts and meaningful evaluation metrics.

Together, these frameworks provide structured methodologies for assessing reliability beyond standard in-distribution performance, offering practical guidance for developing robust and transparent models in high-stakes applications.