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:Isaac Newton Institute Seminar Series
SUMMARY:New directions in solving structured nonconvex pro
blems in multivariate statistics - Rahul Mazumde
r (Massachusetts Institute of Technology)
DTSTART;TZID=Europe/London:20180306T110000
DTEND;TZID=Europe/London:20180306T120000
UID:TALK102301AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/102301
DESCRIPTION:Nonconvex problems arise frequently in modern appl
ied statistics and machine learning\, posing outst
anding challenges from a computational and statist
ical viewpoint. Continuous especially convex optim
ization\, has played a key role in our computation
al understanding of \;(relaxations or approxim
ations of) \; \;these problems. However\,
some other well-grounded techniques in mathematica
l optimization (for example\, mixed integer optimi
zation) have not been explored to their fullest po
tential. When the underlying statistical problem b
ecomes difficult\, simple convex relaxations and/o
r greedy methods have shortcomings. Fortunately\,
many of these can be ameliorated by using estimato
rs that can be posed as solutions to structured di
screte optimization problems. To this end\, I will
demonstrate how techniques in modern computationa
l mathematical optimization (especially\, discrete
optimization) can be used to address the canonica
l problem of best-subset selection and cousins. I
will describe how recent algorithms based on local
combinatorial optimization can lead to high quali
ty solutions in times comparable to (or even faste
r than) the fastest algorithms based on L1-regular
ization. I will also discuss the relatively less u
nderstood low Signal to Noise ratio regime\, where
usual subset selection performs unfavorably from
a \; statistical viewpoint\; and propose simpl
e alternatives that rely on nonconvex optimization
. If time permits\, I will outline problems arisin
g in the context robust statistics (least median s
quares/least trimmed squares)\, low-rank factor an
alysis and nonparametric function estimation where
\, these techniques seem to be promising.

LOCATION:Seminar Room 2\, Newton Institute
CONTACT:info@newton.ac.uk
END:VEVENT
END:VCALENDAR