BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Learning via Data Compression: Bayesian Coresets and Sparse Variat
 ional Inference - Trevor Campbell\, University of British Columbia
DTSTART:20190626T100000Z
DTEND:20190626T110000Z
UID:TALK126718@talks.cam.ac.uk
CONTACT:Robert Peharz
DESCRIPTION:We have reached a point in many fields of science and technolo
 gy where we create data at a pace that far outstrips our capacity to proce
 ss it. While a boon from a statistical perspective\, this wealth of data p
 resents a computational challenge: how might we design a model-based infer
 ence system that learns forever\, retains important past information\, doe
 sn't get bogged down by a persistent stream of new data\, and makes infere
 nces with guaranteed statistical quality? The human nervous system provide
 s inspiration\; to handle the astounding amount of perceptual data it cons
 tantly receives\, the nervous system filters and compresses the data signi
 ficantly before passing it along to the brain where learning occurs. Altho
 ugh a seemingly simple solution\, it does raise interesting questions for 
 the design of a computational inference system: how should we decide what 
 data to retain\, how should we compress it\, and what degree of compressio
 n should we apply before learning from it?  \n\nThis talk will cover recen
 t work on Bayesian coresets ("core of a dataset")\, a methodology for stat
 istical inference via data compression. Coresets achieve compression by fo
 rming a small weighted subset of data that replaces the full dataset durin
 g inference\, leading to significant computational gains with provably min
 imal loss in inferential quality. In particular\, the talk will present nu
 merous methods for Bayesian coreset construction\, from previously-develop
 ed subsampling\, greedy\, and sparse linear regression-based techniques to
  novel algorithms based on sparse variational inference (VI). In contrast 
 to past algorithms\, the sparse VI-based algorithms are fully automated\, 
 requiring only the dataset and probabilistic model specification as inputs
 . The talk will additionally provide a unifying view and statistical analy
 sis of these methods using the theory of exponential families and Riemanni
 an information geometry. The talk will conclude with empirical results sho
 wing that despite requiring much less user input than past methods\, spars
 e VI coreset construction provides state-of-the-art data summarization for
  Bayesian inference. 
LOCATION:Engineering Department\, CBL Room BE-438.
END:VEVENT
END:VCALENDAR