Universität Augsburg
|
Dr. Nikolas Nüsken
Universität Potsdam
spricht am
Donnerstag, 25. Juli 2019
um
15:45 Uhr
im
TUM, Boltzmann-3, Garching, Raum 03.08.011, Etage 3
über das Thema:
Abstract: |
Sampling from high-dimensional probability distributions is an important challenge in computational statistics, molecular dynamics and machine learning. In this talk, I will present two approaches based on interacting particle systems. The first part will exhibit a framework for constructing an ensemble of coupled Markov processes with fixed marginals. As it turns out, many natural design criteria (including the asymptotic variance of estimators) lead to multi-marginal optimal transport problems that can be analysed in terms of infinitesimal generators and solved approximately. In the second part, I will discuss some geometrical ideas in the theory of gradient flows, aimed towards analysing the recently introduced Stein variational gradient descent algorithm. This is joint work with A. Duncan, G. Pavlitiotis and L. Szpruch. |
Hierzu ergeht herzliche Einladung. |