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:Statistics
SUMMARY:Robust inference for intractable likelihood models
using kernel divergences - Francois-Xavier Briol
(UCL)
DTSTART;TZID=Europe/London:20220211T130000
DTEND;TZID=Europe/London:20220211T140000
UID:TALK169907AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/169907
DESCRIPTION:Modern statistics and machine learning tools are b
eing applied to increasingly complex phenomenon\,
and as a result make use of increasingly complex m
odels. A large class of such models are the so-cal
led intractable likelihood models\, where the like
lihood is either too computational expensive to ev
aluate\, or impossible to write down in closed for
m. This creates significant issues for classical a
pproach such as maximum likelihood estimation or B
ayesian inference\, which are entirely reliant on
evaluations of a likelihood. In this talk\, we wil
l cover several novel inference schemes which by-p
ass this issue. These will be constructed from ker
nel-based discrepancies such as maximum mean discr
epancies and kernel Stein discrepancies\, and can
be used either in a frequentist or Bayesian framew
ork. An important feature of our approach is that
it will be provably robust\, in the sense that a s
mall number of outliers or mild model misspecifica
tion will not have a significant impact on paramet
er estimation. In particular\, we will show how th
e choice of kernel can allow us to trade statistic
al efficiency with robustness. The methodology wil
l then be illustrated on a range of intractable li
kelihood models in signal processing and biochemis
try.
LOCATION:MR12\, Centre for Mathematical Sciences
CONTACT:Qingyuan Zhao
END:VEVENT
END:VCALENDAR