IMPAN Colloquium
Wednesdays at 14:15 in Room 321, before the lecture there are cookies and tea in room 409 at 13:45
Organizers:
Marcin Lara, Grigor Sargsyan, Mateusz Wasilewski, Aneta Wróblewska-Kamińska
Embedded Files
UPCOMING MEETINGS
UPCOMING MEETINGS
OCTOBER 8
Błażtej Miasojedow (University of Warsaw)
Sampling as optimisation over the space of measure
Sampling from probability distributions that are only known up to a constant is a key challenge in computational statistics and machine learning. A common method is to use stochastic processes such as Langevin dynamics, which form the basis of many MCMC (Markov chain Monte Carlo) algorithms. In this talk, we present an alternative view: sampling can be seen as an optimisation problem in the space of probability measures equipped with a specific metric, the so-called Wasserstein space W2. From this perspective, sampling algorithms approximate gradient flows in the W2 metric. This approach provides a clearer interpretation of existing methods, extends their use to a wider range of distributions, and allows for more precise convergence guarantees.
Page updated
Google Sites
Report abuse