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:Stability - Bin Yu\, University of California\, Be
rkeley
DTSTART;TZID=Europe/London:20121126T160000
DTEND;TZID=Europe/London:20121126T170000
UID:TALK40802AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/40802
DESCRIPTION:Reproducibility is imperative for any scientific d
iscovery. Often than not\,\nmodern scientific find
ings rely on statistical analysis of high-dimensio
nal\ndata. At a minimum\, reproducibility manifest
s itself in stability of\nstatistical results rela
tive to “reasonable” perturbations to data and to
the\nmodel used. Jacknife\, bootstrap\, and cross-
validation are based on\nperturbations to data\, w
hile robust statistics methods deal with perturbat
ions\nto models.\n\nIn this talk\, a case is made
for the importance of stability in\nstatistics. Fi
rstly\, we motivate the necessity of stability of
interpretable\nencoding\nmodels for movie reconstr
uction from brain fMRI signals. Secondly\, we find
\nstrong evidence in the literature to demonstrate
the central role of stability\nin statistical inf
erence. Thirdly\, a smoothing parameter selector b
ased on\nestimation stability (ES)\, ES-CV\, is pr
oposed for Lasso\, in order to bring\nstability to
bear on cross-validation (CV). ES-CV is then util
ized in the\nencoding models to reduce the number
of predictors by 60% with almost no loss\n(1.3%) o
f prediction performane across over 2\,000 voxels.
Last\, a novel\n“stability” argument is seen to d
rive new results that shed light on the\nintriquin
g interactions between sample to sample varibility
and heavier tail\nerror distribution (e.g. double
-exponential) in high dimensional regression\nmode
ls with p predictors and n independent samples. In
particular\, when p/n →\nκ ∈ (0.3\, 1) and error
is double-exponential\, OLS is a better estimator
than\nLAD.
LOCATION:MR12\, CMS\, Wilberforce Road\, Cambridge\, CB3 0W
B
CONTACT:Richard Samworth
END:VEVENT
END:VCALENDAR