- The Statistics group of the Mathematical Institute meets weekly on Mondays, 13.00-13.45.
- Due to the Corona virus, the seminar is held online for an indefinite time.
23-11-2020 | Stefan Franssen
Frequentist coverage of empirical Bayesian uncertainty quantification using Deep Neural Network regression in Besov spaces
In the past 5 years there has been a breakthrough in our understanding of the behaviour of (sparse) Deep Neural network regression. For $\beta$-Hölder spaces, Johannes gave near minimax convergence rates, and the work has been extended to Besov spaces by Suzuki. These works give guarantees for the square loss of (near) minimizers of the empirical square loss, which imply that Deep Neural Networks following their designs will have good uncertainty quantification. In spite of this progress, there has not been any rigorous way of quantifying uncertainty in the estimates of Deep Neural Networks. We provide both an Empirical Bayesian methodology to provide uncertainty quantification and a theoretical analysis with frequentist coverage guarantees. We also ran a simulation study which illustrates the coverage properties. In this talk, we will take a closer look at the methodology, illustrated with simulation study pictures and a short sketch of the theoretical results.
30-11-2020 | Jacqueline Meulman
07-12-2020 | George Kantidakis
14-12-2020 | Johannes Schmidt-Hieber
11-01-2021 | Lasse Vuursteen
18-01-2021 | Stéphanie van der Pas
25-01-2021 | Marten Kampert
01-02-2021 | Geerten Koers
08-02-2021 | Valentina Masarotto
15-02-2021 | Laura Zwep
22-02-2021 | Bart Eggen
01-02-2021 | Marta Spreafico
08-02-2021 | Thijs Bos
15-02-2020 | Aad van der Vaart