We cordially invite you to the colloquium of Lars Doorenbos!
On September 12, 10:00 to 11:00 a.m, Lars Doorenbos, Doktoral Student at the University of Bern (Schwitzerland), will give a research talk in English.
The research talk is dedicated to the topic “Reliably Processing Unexpected and Ambiguous Images in Computer Vision“.
Abstract:
Artificial intelligence (AI) has gone beyond an academic curiosity and now has a real impact on everyday life. As the rapid advancements in AI continue, questions regarding the reliability of these systems become increasingly important. Handling uncertainty, in particular, is a crucial challenge for deploying these systems in real-world environments, as evaluating unexpected or ambiguous samples comes with few guarantees, and performance is typically poorer than expected. These shortcomings limit the usability of machine learning in safety-critical applications. In this talk, I will focus on two aspects of handling uncertainty. First, I will present our research on out-of-distribution detection through the lens of data invariants, which can filter out unexpected samples with high epistemic uncertainty. Second, I will show our work on aleatoric uncertainty modeling with categorical diffusion models, which are capable of representing the inherent ambiguity present in images. Together, these techniques allow for a comprehensive view of uncertainty, facilitating the development of reliable, real-world machine learning systems.
The event is free of charge. Interested individuals are warmly invited to attend!