BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Machine Learning @ CUED
SUMMARY:Learning via Data Compression: Bayesian Coresets a
nd Sparse Variational Inference - Trevor Campbell\
, University of British Columbia
DTSTART;TZID=Europe/London:20190626T110000
DTEND;TZID=Europe/London:20190626T120000
UID:TALK126718AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/126718
DESCRIPTION:We have reached a point in many fields of science
and technology where we create data at a pace that
far outstrips our capacity to process it. While a
boon from a statistical perspective\, this wealth
of data presents a computational challenge: how m
ight we design a model-based inference system that
learns forever\, retains important past informati
on\, doesn't get bogged down by a persistent strea
m of new data\, and makes inferences with guarante
ed statistical quality? The human nervous system p
rovides inspiration\; to handle the astounding amo
unt of perceptual data it constantly receives\, th
e nervous system filters and compresses the data s
ignificantly before passing it along to the brain
where learning occurs. Although a seemingly simple
solution\, it does raise interesting questions fo
r the design of a computational inference system:
how should we decide what data to retain\, how sho
uld we compress it\, and what degree of compressio
n should we apply before learning from it? \n\nTh
is talk will cover recent work on Bayesian coreset
s ("core of a dataset")\, a methodology for statis
tical inference via data compression. Coresets ach
ieve compression by forming a small weighted subse
t of data that replaces the full dataset during in
ference\, leading to significant computational gai
ns with provably minimal loss in inferential quali
ty. In particular\, the talk will present numerous
methods for Bayesian coreset construction\, from
previously-developed subsampling\, greedy\, and sp
arse linear regression-based techniques to novel a
lgorithms based on sparse variational inference (V
I). In contrast to past algorithms\, the sparse VI
-based algorithms are fully automated\, requiring
only the dataset and probabilistic model specifica
tion as inputs. The talk will additionally provide
a unifying view and statistical analysis of these
methods using the theory of exponential families
and Riemannian information geometry. The talk will
conclude with empirical results showing that desp
ite requiring much less user input than past metho
ds\, sparse VI coreset construction provides state
-of-the-art data summarization for Bayesian infere
nce.
LOCATION:Engineering Department\, CBL Room BE-438.
CONTACT:Robert Peharz
END:VEVENT
END:VCALENDAR