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:Machine Learning @ CUED
SUMMARY:Nonparametric Generative Modeling via Optimal Tran
sport and Diffusions with Provable Guarantees - Um
ut Şimşekli\, Télécom Paristech
DTSTART;TZID=Europe/London:20190507T110000
DTEND;TZID=Europe/London:20190507T120000
UID:TALK122029AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/122029
DESCRIPTION:By building up on the recent theory that establish
ed the connection between implicit generative mode
ling and optimal transport\, in this talk\, I will
present a novel parameter-free algorithm for lear
ning the underlying distributions of complicated d
atasets and sampling from them. The proposed algor
ithm is based on a functional optimization problem
\, which aims at finding a measure that is 'close
to the data distribution as much as possible' and
also 'expressive enough' for generative modeling p
urposes. The problem will be formulated as a gradi
ent flow in the space of probability measures. The
connections between gradient flows and stochastic
differential equations will let us develop a comp
utationally efficient algorithm for solving the op
timization problem\, where the resulting algorithm
will resemble the recent dynamics-based Markov Ch
ain Monte Carlo algorithms. I will then present fi
nite-time error guarantees for the proposed algori
thm. I will finally present some experimental resu
lts\, which support our theory and shows that our
algorithm is able to capture the structure of chal
lenging distributions. \n\nIf time permits\, I wil
l also talk about possible extensions of this appr
oach.\n\nThe talk will be based on these two artic
les: \n1) "Sliced-Wasserstein Flows":https://arxiv
.org/abs/1806.08141\n2) "Generalized Sliced Wasser
stein Distances":https://arxiv.org/abs/1902.00434
LOCATION:Engineering Department\, CBL Room BE-438.
CONTACT:Eric T Nalisnick
END:VEVENT
END:VCALENDAR