BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Generalized Sliced-Wasserstein Distances - Soheil Kouri\, HRL Labo
 ratories
DTSTART:20190423T140000Z
DTEND:20190423T150000Z
UID:TALK120643@talks.cam.ac.uk
CONTACT:Matthew Thorpe
DESCRIPTION:Emerging from the optimal transportation problem and due to th
 eir favorable geometric properties\, Wasserstein distances have recently a
 ttracted ample attention from the machine learning and signal processing c
 ommunities. Wasserstein distances have been used in supervised\, semi-supe
 rvised\, and unsupervised learning problems\, as well as in domain adaptat
 ion and transfer learning. However\, the application of Wasserstein distan
 ces to high-dimensional probability measures is often hindered by their ex
 pensive computational cost. Sliced-Wasserstein (SW) distances\, on the oth
 er hand\, have similar qualitative properties to the Wasserstein distances
  but are significantly simpler to compute. The simplicity of computation o
 f this distance has motivated recent work to use SW as a substitute for th
 e Wasserstein distances. In this presentation\, I first review the mathema
 tical concepts behind sliced Wasserstein distances. Then I introduce an en
 tire class of new distances\, denoted as Generalized Sliced-Wasserstein (G
 SW) distances\, that extends the idea of linear slicing used in SW distanc
 es to general non-linear slicing of probability measures. Finally\, I will
  review various applications of SW and GSW in deep generative modeling and
  transfer learning.
LOCATION:MR 14
END:VEVENT
END:VCALENDAR