BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Probabilistic computing: computation as universal stochastic infer
 ence\, not deterministic calculation - Vikash K. Mansinghka (MIT)
DTSTART:20120223T123000Z
DTEND:20120223T133000Z
UID:TALK36623@talks.cam.ac.uk
CONTACT:Zoubin Ghahramani
DESCRIPTION:Latent variable modeling and Bayesian inference are appealing 
 in\ntheory --- they provide a unified mathematical framework for solving a
 \nwide range of machine learning problems --- but are often difficult to\n
 apply effectively in practice. Accurate inference in even simple\nmodels c
 an seem computationally intractable\, while more realistic\nmodels are dif
 ficult to even write down precisely.\n\nIn this talk\, I will introduce ne
 w probabilistic programming\ntechnology that aims to alleviave these diffi
 culties. Unlike\ngraphical models\, which marries statistics with graph th
 eory\,\nprobabilistic programming marries Bayesian inference with universa
 l\ncomputation. Probabilistic programming can make it easier to build\nuse
 ful\, fast machine learning software that goes significantly beyond\ngraph
 ical models in flexibility and power. I will illustrate\nprobabilistic pro
 gramming using page-long probabilistic programs that\nbreak simple CAPTCHA
 s --- by running randomized CAPTCHA generators backwards --- interpret noi
 sy time-series data from clinical medicine\, and estimate good predictive 
 models for arbitrary structured data tables without any parameter tuning o
 r pre-processing.\n\nI will also describe stochastic digital circuit archi
 tectures that carry these principles down to the physical layer and yield 
 1000x speed and 10-100x power improvements over deterministic designs on p
 roblems of optical flow\, clustering\, and inference in discrete graphical
  models.\n\nThroughout\, I will highlight the ways probabilistic programmi
 ng points\nthe way to a new model of computation\, based on universal infe
 rence\nover distributions rather than universal calculation of functions\,
  and\nexposes the mathematical and algorithmic structure needed to enginee
 r efficient\, distributed machine learning systems.
LOCATION:Engineering Department\, CBL Room BE-438
END:VEVENT
END:VCALENDAR