BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Bayesian optimisation in many dimensions with bespoke probabilisti
 c programs - Valentin Dalibart
DTSTART:20170210T110000Z
DTEND:20170210T120000Z
UID:TALK71048@talks.cam.ac.uk
CONTACT:39846
DESCRIPTION:In this talk\, I will present a collection of techniques to ma
 ke Bayesian optimisation converge in grey-box optimisation problems with m
 any dimensions. First\, I will discuss how better priors of the objective 
 function can lead to orders of magnitude improvements in convergence. I wi
 ll use probabilistic programming to build probabilistic models that have g
 ood convergence and can leverage many observable properties of the objecti
 ve function for inference. I will introduce a class of probabilistic progr
 ams that are both useful for Bayesian optimisation and support inference a
 t a reasonable computational cost. Second\, I will discuss techniques to h
 elp the numerical optimisation stage of the Bayesian optimisation converge
 \, when algorithms such as DIRECT or CMA-ES are not sufficient.\n\nI will 
 present applications of these techniques to optimise the configuration of 
 computer systems\, such as TensorFlow\, and maximise their computational p
 erformance. The techniques will be exemplified using the BOAT framework (a
  framework to build BespOke Auto-Tuners) which is open source and availabl
 e at https://github.com/VDalibard/BOAT.
LOCATION:CBL Room BE-438\, Department of Engineering
END:VEVENT
END:VCALENDAR