We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: Sampling From the Wasserstein Barycenter

Authors: Chiheb Daaloul (1), Thibaut Le Gouic (2), Jacques Liandrat (1), Magali Tournus (1) ((1) Aix-Marseille Univ., CNRS, I2M, UMR7373, Centrale Marseille, Marseille, France, (2) Massachusetts Institute of Technology, Department of Mathematics, USA)
Abstract: This work presents an algorithm to sample from the Wasserstein barycenter of absolutely continuous measures. Our method is based on the gradient flow of the multimarginal formulation of the Wasserstein barycenter, with an additive penalization to account for the marginal constraints. We prove that the minimum of this penalized multimarginal formulation is achieved for a coupling that is close to the Wasserstein barycenter. The performances of the algorithm are showcased in several settings.
Subjects: Machine Learning (cs.LG); Machine Learning (stat.ML)
Cite as: arXiv:2105.01706 [cs.LG]
  (or arXiv:2105.01706v1 [cs.LG] for this version)

Submission history

From: Chiheb Daaloul [view email]
[v1] Tue, 4 May 2021 18:57:41 GMT (1694kb,D)

Link back to: arXiv, form interface, contact.