Research Group "Stochastic Algorithms and Nonparametric Statistics"
Seminar "Modern Methods in Applied Stochastics and Nonparametric Statistics" Summer Semester 2026
|
|
| 14.04.2026 | |
| 21.04.2026 | |
| |
|
| 28.04.2026 | |
| 05.05.2026 | |
| | |
| 12.05.2026 | Dr. Pavel Dvurechensky (WIAS Berlin) |
| On global rates for semismooth Newton method | |
| 19.05.2026 | Sorelle Toukam (WIAS-Berlin) |
| Stochastic maximum principle for McKean-Vlasov SDEs with rough drift coefficients
|
|
| 26.05.2026 | |
| |
|
| 02.06.2026 | |
| ESH, AWA-Str. 39 | |
| 09.06.2026 | |
| |
|
| 16.06.2026 | Jakob Kellermann (WIAS Berlin) |
| Signature kernel ridge regression
|
|
| 23.06.2026 | WIAS excursion |
|
|
|
| 30.06.2026 | Oleg Butkovsky (WIAS Berlin) |
| Will there still be mathematicians in 1 year? A personal experience from the 'The First Proof' project
|
|
| 07.07.2026 | Jeyong Lee (WIAS Berlin) |
| HVP 11 a, R. 313 | Bayesian online learning Bayesian online learning provides a coherent framework for sequential inference, in which posterior beliefs are updated as new data become available. However, recursive Bayesian updating is rarely computationally tractable outside conjugate models, and posterior approximation is often required. In this talk, we present our recent work on Bayesian online learning. We first show that sequential variational approximation can be asymptotically valid: once the mini-batch size exceeds a threshold depending on the parameter dimension, the accumulated approximation error becomes negligible, and the final online posterior is asymptotically indistinguishable from the full posterior. We then consider the one-pass setting and propose a new Bayesian online learning algorithm, showing that valid statistical inference remains possible without requiring the mini-batch size to diverge. Together, these results provide a theoretical foundation for valid uncertainty quantification in Bayesian online learning. |
| 14.07.2026 | |
| |
|
| 21.07.2026 | |
| |
last reviewed: July 7, 2026 by Christine Schneider