Title: From Test-Time Learning to Uncertainty-Aware Medical Foundation Models

Abstract: Modern medical AI systems are increasingly deployed in dynamic clinical environments where data distributions continuously shift across hospitals, scanners, patient populations, and evolving clinical practices. His talk will focus on test-time learning methods that enable models to adapt after deployment without requiring retraining or access to source data. He will further present recent work on uncertainty-aware medical foundation models, including conformal prediction techniques that provide rigorous confidence guarantees under distribution shift. Together, these approaches offer a path toward safer, more reliable, and clinically deployable AI systems for healthcare.

Biography: Associate Professor at Aarhus University and Founder and Director of the Adaptive and Agentic AI Lab (A3 Lab). He is an ELLIS Scholar, an AIAS Associate Fellow, and is affiliated with the Pioneer Centre for AI.His research focuses on trustworthy and adaptive AI for safety-critical applications, spanning test-time adaptation and updates, uncertainty quantification, multimodal foundation models, and medical image analysis. He has made contributions to self-supervised learning, open-set recognition, continual adaptation, and robust deployment of AI systems under distribution shift. He is the organizer of the 3rd Workshop on Test-Time Updates (TTU) at ICLR 2026 and serves in senior reviewing and area-chair roles for leading AI conferences, including NeurIPS, ICLR, ICML, CVPR, NeurIPS, and AAAI.