BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Scaling and Generalizing Approximate Bayesian Inference - Prof. Da
 vid Blei (Columbia University)
DTSTART:20160712T100000Z
DTEND:20160712T110000Z
UID:TALK66805@talks.cam.ac.uk
CONTACT:Louise Segar
DESCRIPTION:Latent variable models have become a key tool for the modern\n
 statistician\, letting us express complex assumptions about the hidden\nst
 ructures that underlie our data. Latent variable models have been\nsuccess
 fully applied in numerous fields.\n\nThe central computational problem in 
 latent variable modeling is\nposterior inference\, the problem of approxim
 ating the conditional\ndistribution of the latent variables given the obse
 rvations.\nPosterior inference is central to both exploratory tasks and\np
 redictive tasks.  Approximate posterior inference algorithms have\nrevolut
 ionized Bayesian statistics\, revealing its potential as a\nusable and gen
 eral-purpose language for data analysis.\n\nBayesian statistics\, however\
 , has not yet reached this potential.\nFirst\, statisticians and scientist
 s regularly encounter massive data\nsets\, but existing approximate infere
 nce algorithms do not scale well.\nSecond\, most approximate inference alg
 orithms are not generic\; each\nmust be adapted to the specific model at h
 and.\n\nIn this talk I will discuss our recent research on addressing thes
 e\ntwo limitations.  I will describe stochastic variational inference\, an
 \napproximate inference algorithm for handling massive data sets.  I\nwill
  demonstrate its application to probabilistic topic models of text\ncondit
 ioned on millions of articles. Then I will discuss black box\nvariational 
 inference.  Black box inference is a generic algorithm for\napproximating 
 the posterior.  We can easily apply it to many models\nwith little model-s
 pecific derivation and few restrictions on their\nproperties.  I will demo
 nstrate its use on longitudinal models of\nhealthcare data\, deep exponent
 ial families\, and discuss a new\nblack-box variational inference algorith
 m in the Stan programming\nlanguage.\n\nThis is joint work based on these 
 three papers:\n\nM. Hoffman\, D. Blei\, J. Paisley\, and C. Wang.  Stochas
 tic variational\ninference.  Journal of Machine Learning Research\, 14:130
 3-1347\, 2013.\n\n   http://www.cs.columbia.edu/~blei/papers/HoffmanBleiWa
 ngPaisley2013.pdf\n\nR. Ranganath\, S. Gerrish\, and D. Blei.  Black box v
 ariational\ninference.  Artificial Intelligence and Statistics\, 2014.\n\n
    http://www.cs.columbia.edu/~blei/papers/RanganathGerrishBlei2014.pdf\n\
 nA. Kucukelbir\, R. Ranganath\, A. Gelman\, and D. Blei.  Automatic\nvaria
 tional inference in Stan.  Neural Information Processing Systems\,\n2015.\
 n\n   http://www.cs.columbia.edu/~blei/papers/KucukelbirRanganathGelmanBle
 i2015.pdf\n
LOCATION:James Dyson Building Meeting Room on the Ground Floor
END:VEVENT
END:VCALENDAR