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:MRC Biostatistics Unit Seminars
SUMMARY:BSU Seminar: "\;Sparse Hamiltonian Flows (or:
Bayesian Coresets Without all the Fuss)"\; - P
rof Trevor Campbell\, The University of British Co
lumbia
DTSTART;TZID=Europe/London:20220426T140000
DTEND;TZID=Europe/London:20220426T150000
UID:TALK172058AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/172058
DESCRIPTION:Bayesian inference provides a coherent approach to
learning from data and uncertainty assessment in
complex\, expressive statistical models. However\,
algorithms for performing inference have not yet
caught up to the deluge of data in modern applicat
ions. One approach---Bayesian coresets---involves
replacing the large dataset with a small\, weighte
d\, representative subset of data during inference
. The coreset is designed to capture the informati
on from the full dataset\, but be much less comput
ationally expensive to store in memory and iterate
over. Although the methodology is sound in princi
ple\, efficiently constructing such a coreset in p
ractice remains a significant challenge: current m
ethods tend to be complicated to implement\, slow\
, require a secondary inference step after coreset
construction\, and do not enable model selection.
In this talk\, I will introduce a new method---sp
arse Hamiltonian flows---that addresses all of the
se challenges. The method involves first subsampli
ng the data uniformly\, and then optimizing a Hami
ltonian flow parametrized by coreset weights and i
ncluding periodic momentum quasi-refreshment steps
. I will present theoretical results demonstrating
that the method enables an exponential compressio
n of the dataset in representative models\, and th
at the quasi-refreshment steps reduce the KL diver
gence to the target. Real and synthetic experiment
s demonstrate that sparse Hamiltonian flows provid
e accurate posterior approximations with significa
ntly reduced runtime compared with competing dynam
ical-system-based inference methods.\n\nThis talk
will be based on two papers that are available onl
ine as preprints:\nChen et al\, "Bayesian coresets
via sparse Hamiltonian flows\," https://arxiv.org
/abs/2203.05723\nNaik et al\, "Fast Bayesian cores
ets via subsampling and quasi-Newton refinement\,"
https://arxiv.org/abs/2203.09675\n
LOCATION:This will be a virtual seminar. FREE registration:
https://us02web.zoom.us/meeting/register/tZ0rc-2v
pz4qHNxdJ4Cm4hyLEn4BVOG5YcIt
CONTACT:Alison Quenault
END:VEVENT
END:VCALENDAR