Humboldt-Universität zu Berlin - Mathematisch-Naturwissenschaftliche Fakultät - Institut für Informatik

Humboldt-Universität zu Berlin | Mathematisch-Naturwissenschaftliche Fakultät | Institut für Informatik | Institutstermine | Kolloquiumsvortrag: "Scalable, Calibrated (Deep) Probabilistic Learning"

Kolloquiumsvortrag: "Scalable, Calibrated (Deep) Probabilistic Learning"

Prof. Dr. Nadja Klein (HU Berlin, Lehrstuhl für Statistik und Data Science, Wirtschaftswissenschaftliche Fakultät) zum Thema "Scalable, Calibrated (Deep) Probabilistic Learning"

Der Kolloquiumsvortrag von findet am Mittwoch, den 4.5.22 ab 16h00 (s.t.) als Zoom-Meeting statt. Eine Zoom-Einladung finden Sie hier. (nur mit Infromatik-Account)



Extracting knowledge from large data sources (Big Data) poses great challenges and chances for statistical methodological research and its interdisciplinary applications. In this talk, we are in the context of probabilistic supervised learning where so-called distributional regression models have gained considerable interest in the statistical literature over the last decade. Distributional regression models are a popular extension to mean regression models that aim at estimating the entire conditional distribution of a response rather than just the mean. However, large-scale data, complex interactions between features and outputs and highly-parameterized models make classical approaches and inferential schemes insufficient or infeasible for both theoretical and computational reasons. In this talk, we will first sketch the idea and flexibility of Bayesian structured additive distributional regression and its estimation. Resulting limitations will then motivate a second branch of distributional regression models based on an implicit copula construction and scalable inference using variational Bayes. We illustrate this approach along an example for marginally calibrated response distributions for end-to-end learning in autonomous driving based on a large amount of image/video data. Given the high safety requirements in this field, accurate uncertainty quantification and proper calibration are crucial.