Abstract

The field of probabilistic circuits has seen a strong development during the last decade, in part by riding on the coattails of the deep learning colossus. The idea of building an uncertainty representation in the form of a computation graph with simple components has been shown very successful for classical probability. The models built and techniques used can serve as inspiration for researchers in the imprecise probability community, and this is my goal for this talk. It will consist of three parts. First, I will discuss (imprecise) probabilistic circuits from a conceptual viewpoint, so that we can dream up many variants without being influenced too much by concrete model class definitions and computational constraints. Then, I will present results of two research lines at the Uncertainty in AI research group of the TU/e. Namely, I'll focus on credal sum-product networks as a concrete imprecise probabilistic model class and on probabilistic integral circuits as the current state of the art for classical probability enabled by deep learning techniques

Bio

Erik Quaeghebeur started his academic career in Gert de Cooman's group at Ghent University working on quite foundation-oriented work in imprecise probability theory, occupying himself with lower previsions and desirable gambles. He was set on continuing this line of work and did so in a number of nice places in Europe and across the Atlantic. He then somewhat surprised himself by moving to wind energy research and strengthening his application-oriented research skills. Afterwards, he joined Cassio de Campos's group at Eindhoven University of Technology and had the opportunity to diversify in other ways as well, collaborating on research in deep learning, uncertainty quantification for mechanical engineering, and also back to foundational work in the field of imprecise probabilities. Since he started in academia, he has always been involved in teaching, from instruction sessions on probability theory over systems & signals courses to a foundations of AI course and a course on, essentially, generalizations of probability theory. Since a couple of years, he has entered education management as well, becoming responsible for a large master's program on Data Science & AI. He enjoys talking about research, teaching, and the other things in life.

Host

Alessandro Facchini,  Associate Professor of Epistemology, Logic and Ethics of Artificial Intelligence, and Co-Head of the Bachelor's Degree Programme in Data Science and Artificial Intelligence at SUPSI.