BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Talks.cam//talks.cam.ac.uk// X-WR-CALNAME:Talks.cam BEGIN:VEVENT SUMMARY:Bayesian Optimization for Probabilistic Programs - Tom Rainforth\, University of Oxford DTSTART:20170315T110000Z DTEND:20170315T120000Z UID:TALK71612@talks.cam.ac.uk CONTACT:12852 DESCRIPTION:Probabilistic programming systems (PPS) allow probabilistic mo dels to be represented in the form of a generative model and statements fo r conditioning on data. Their core philosophy is to decouple model specifi cation and inference\, the former corresponding to the user-specified prog ram code and the latter to an inference engine capable of operating on arb itrary programs. Removing the need for users to write inference algorithm s significantly reduces the burden of developing new models and makes effe ctive statistical methods accessible to non-experts.\n\nIn this talk I wil l present BOPP: a general purpose framework for (type II) maximum likeliho od and marginal maximum a posteriori estimation of probabilistic program v ariables. By using a series of code transformations\, the evidence of any probabilistic program\, and therefore of any graphical model\, can be opti mized with respect to an arbitrary subset of its sampled variables. To ca rry out this optimization\, we develop the first Bayesian optimization pac kage to directly exploit the source code of its target\, leading to innova tions in problem-independent hyperpriors\, unbounded optimization\, and im plicit constraint satisfaction\; delivering significant performance improv ements over prominent existing packages. We present applications of our m ethod to a number of tasks including engineering design and parameter opti mization.\n\n3 minute version of the talk: http://youtu.be/gVzV-NxKa9U\n\n Associated paper: http://papers.nips.cc/paper/6421-bayesian-optimization-f or-probabilistic-programs\n\nCode: http://github.com/probprog/bopp LOCATION: Cambridge University Engineering Department\, CBL Seminar room B E4-38 END:VEVENT END:VCALENDAR