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:Bayesian optimisation in many dimensions with besp
oke probabilistic programs - Valentin Dalibart
DTSTART;TZID=Europe/London:20170210T110000
DTEND;TZID=Europe/London:20170210T120000
UID:TALK71048AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/71048
DESCRIPTION:In this talk\, I will present a collection of tech
niques to make Bayesian optimisation converge in g
rey-box optimisation problems with many dimensions
. First\, I will discuss how better priors of the
objective function can lead to orders of magnitude
improvements in convergence. I will use probabili
stic programming to build probabilistic models tha
t have good convergence and can leverage many obse
rvable properties of the objective function for in
ference. I will introduce a class of probabilistic
programs that are both useful for Bayesian optimi
sation and support inference at a reasonable compu
tational cost. Second\, I will discuss techniques
to help the numerical optimisation stage of the Ba
yesian optimisation converge\, when algorithms suc
h as DIRECT or CMA-ES are not sufficient.\n\nI wil
l present applications of these techniques to opti
mise the configuration of computer systems\, such
as TensorFlow\, and maximise their computational p
erformance. The techniques will be exemplified usi
ng the BOAT framework (a framework to build BespOk
e Auto-Tuners) which is open source and available
at https://github.com/VDalibard/BOAT.
LOCATION:CBL Room BE-438\, Department of Engineering
CONTACT:
END:VEVENT
END:VCALENDAR