Abstract

Probability is often treated as a technical tool for statistics. In this talk, I will argue for a broader perspective: probability is best understood as a language for rational reasoning under uncertainty — the natural counterpart of logic. While logic tells us what follows from facts that are certainly true, probability allows us to reason when knowledge is incomplete, noisy, or ambiguous. Starting from this viewpoint, I will give a personal tour through my research agenda in probabilistic machine learning. Since my PhD, I have been fascinated by the question of how probabilistic reasoning can be made practical for modern AI systems. A central challenge is that exact probabilistic inference is usually computationally intractable. Much of my work therefore focuses on “tractable” probabilistic models, in particular probabilistic circuits, which make complex reasoning feasible and turn probability into a practical tool rather than a purely theoretical ideal. At the same time, I will show how these foundational ideas connect naturally to applications in AI for Science. Here, uncertainty is not a nuisance but often the key ingredient: from designing new materials and molecules to optimizing experiments, probabilistic machine learning provides principled ways to learn from scarce data and to decide what to try next. Bayesian optimization, one of the methods I use frequently, offers a particularly elegant example of this interplay between reasoning, uncertainty, and scientific discovery.