<ul><li>This thesis focuses on the reliability and safety of machine learning models in open-world deployment by addressing distributional uncertainty and unknown classes.</li><li>Novel frameworks are introduced to optimize for in-distribution accuracy and reliability to unseen data, including an unknown-aware learning framework.</li><li>Outlier synthesis methods like VOS, NPOS, and DREAM-OOD are proposed to generate informative unknowns during training, enhancing OOD detection using unlabeled data.</li><li>The thesis extends reliable learning to foundation models through tools like HaloScope, MLLMGuard, and data cleaning methods, aiming to improve the safety of large-scale models in deployment.</li></ul>

Foundations of Unknown-aware Machine Learning

Discover more