Date and time: Monday 14 September 2026, 11:00-12:00 CEST
Speaker: Vincent Fortuin, University of Technology Nuremberg
Title: Uncertainty-aware decision making with Bayesian deep learning
Where: Fantum, Lindstedtsvägen 24, KTH main campus
Directions: https://www.kth.se/places/room/id/c9ec01ab-b536-4be6-b82a-0d52ddadb2e6
Host: Martin Trapp mtrapp@kth.se
Bio: Vincent Fortuin is a full professor of probabilistic machine learning at the University of Technology Nuremberg and a research group leader at Helmholtz AI in Munich, leading the group for Efficient Learning and Probabilistic Inference for Science (ELPIS). He is also a Branco Weiss Fellow, an ELLIS Scholar, a Fellow of the Konrad Zuse School of Excellence in Reliable AI, and a Senior Researcher at the Munich Center for Machine Learning.
His research focuses on reliable and data-efficient AI approaches, especially for scientific applications, leveraging Bayesian deep learning, sequential decision making, and amortized inference. Before that, he did his PhD in Machine Learning at ETH Zürich and was a Research Fellow at the University of Cambridge. He is also a board member of the ISBA section for Bayesian AI and the Symposium on Probabilistic Machine Learning (ProbML).
Website: https://fortuin.github.io
Abstract: While traditional benchmarking in machine learning has often focused on the predictive accuracy alone, in this day and age of agentic AI, it becomes increasingly crucial to think about downstream decision outcomes. In this talk, I will question how well conventional metrics align with decision outcomes [spoiler: not very well] and propose a new class of decision-aligned metrics. I will also show recent results in Bayesian deep learning, supporting the view that well-specified weight-space priors for Bayesian neural networks exist and are learnable. I will then close by showing how using Bayesian deep learning can improve decision outcomes in real-world sequential learning tasks.
