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:CMI Student Seminar Series
SUMMARY:Dimension-Robust Function Space MCMC With Neural N
etwork Priors - Torben Sell (University of Cambrid
ge)
DTSTART;TZID=Europe/London:20210414T140000
DTEND;TZID=Europe/London:20210414T150000
UID:TALK158176AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/158176
DESCRIPTION:At the beginning of this talk\, two popular priors
defined on function spaces are discussed: Gaussia
n priors\, which come with a set of orthogonal bas
is functions\, and Bayesian Neural Networks (BNNs)
\, which are popular in the machine learning commu
nity. I argue that both priors come with disadvant
ages\, and propose a new class of BNN priors that
alleviate them. The resulting posteriors are amena
ble to sampling using Hilbert space Markov chain M
onte Carlo methods (unlike standard BNNs)\, and sc
ale more favourably in the dimension of the functi
on’s domain (unlike most Gaussian measures). Some
theoretical results as well as numerical illustrat
ions are presented\, and my talk will end by posin
g future research directions. This talk is loosely
based on the following preprint: https://arxiv.or
g/abs/2012.10943.
LOCATION:https://maths-cam-ac-uk.zoom.us/j/95531783868?pwd=
U3pPbmYxTXZYRVZMWFBVTkVnWmUvZz09
CONTACT:Neil Deo
END:VEVENT
END:VCALENDAR