Probabilistic Machine Learning

Prof. Dr. Vincent Fortuin

Deep learning achieves impressive results, but can be overly confident and usually requires large training datasets. However, data is scarce in many scientific applications, while extensive prior knowledge is available. Here, insights from Bayesian statistics can help to drastically improve these models.

We maintain that the goal of machine learning should not exclusively be to make better predictions, but to lead to better decision outcomes. Uncertainty estimation and probabilistic approaches are crucial to this. Important research questions in this context include

how to effectively specify priors in deep Bayesian models,
how to harness unlabeled data to learn reusable representations,
how to transfer knowledge between tasks using meta-learning, and
how to guarantee generalization performance using PAC-Bayesian bounds.

The mission statement of our research group is to make Bayesian deep learning a viable standard solution for scientific learning tasks. To this end, we combine basic research on priors and inference with applied research to make methods more practical. In close collaboration with scientists, we anchor our approaches in real-world problems.

Bayesian thinking helps us update our own beliefs in light of new evidence. Curiosity motivates us to really get to the core of questions. Hope guides us—both regarding both regarding the success of each new research project and the overall impact that our research will have on the world.