BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//UNL_UCBCN//NONSGML UNL Event Publisher//EN
X-WR-CALNAME:Environmental Studies
CALSCALE:GREGORIAN
METHOD:PUBLISH
BEGIN:VEVENT
DTSTART:20250221T220000Z
UID:186731@events.unl.edu
DTSTAMP:20250114T180403Z
ORGANIZER;CN=Jack Jeffries:
SUMMARY:Shiying Li
STATUS:CONFIRMED
DESCRIPTION:Matching and estimating probability distributions are fundament
 al tasks in generative modeling applications. While optimal transport (OT)
  provides a natural framework for matching distributions\, its large-scale
  computation remains challenging\, especially in high dimensions. Our focu
 s lies on an efficient iterative slice-matching approach initially introdu
 ced by Pitie et al. for transferring color statistics. The resulting slice
 d optimal transport is constructed by solving a sequence of one-dimensiona
 l OT problems between the sliced distributions\, which are fast to compute
  and admit closed-form solutions. While these techniques have proven effec
 tive in various data science applications\, a comprehensive understanding 
 of their convergence properties remains lacking. Our primary aim is to est
 ablish an almost sure convergence\, shedding light on the connections with
  stochastic gradient descent in the context of the sliced-Wasserstein dist
 ance. If time permits\, we will also explore invariance\, equivariance\, a
 nd Lipschitz properties associated with one-step of the procedure\, yieldi
 ng recovery and stability results for sliced OT approximations with a sing
 le step. This talk is based on joint work with Caroline Moosmueller.
LOCATION:Avery Hall Room 115
URL://events.unl.edu/ENVR/2025/02/21/186731/
DTEND:20250221T225000Z
END:VEVENT
END:VCALENDAR