BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:CMI Student Seminar Series
SUMMARY:Stochastic Optimization for Wasserstein Estimators
- Marin Ballu (University of Cambridge)
DTSTART;TZID=Europe/London:20210609T140000
DTEND;TZID=Europe/London:20210609T150000
UID:TALK159952AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/159952
DESCRIPTION:Optimal transport is a foundational problem in\nop
timization\, that allows to compare probability\nd
istributions while taking into account geometric\n
aspects. Its optimal objective value\, the Wassers
tein distance\, provides an important loss between
\ndistributions that has been used in many applica
tions throughout machine learning and statistics.\
nRecent algorithmic progress on this problem and\n
its regularized versions have made these tools inc
reasingly popular. However\, existing techniques (
pre-2020)\nrequire solving an optimization problem
to obtain a single gradient of the loss\, thus sl
owing\ndown first-order methods to minimize the su
m of\nlosses\, that require many such gradient com
putations. In this talk\, I will introduce an algo
rithm\nto solve a regularized version of this prob
lem\nof Wasserstein estimators\, with a time per s
tep\nwhich is sublinear in the natural dimensions
of\nthe problem.
LOCATION:https://maths-cam-ac-uk.zoom.us/j/95531783868?pwd=
U3pPbmYxTXZYRVZMWFBVTkVnWmUvZz09
CONTACT:Neil Deo
END:VEVENT
END:VCALENDAR