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:Deep Gaussian processes and variational propagatio
n of uncertainty - Andreas Damianou - Sheffield Un
iversity
DTSTART;TZID=Europe/London:20150629T110000
DTEND;TZID=Europe/London:20150629T120000
UID:TALK59975AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/59975
DESCRIPTION:The complex and potentially high-dimensional natur
e of real world data renders them difficult to vis
ualise\, understand and predict. This talk will di
scuss a family of Bayesian approaches for solving
the aforementioned problems by combining the flexi
bility of Gaussian processes with the expressivene
ss of graphical and latent variable models. The ge
neral framework is referred to as a deep Gaussian
process\, from which interesting special cases eme
rge\, for example time-series and multi-view model
s. The framework is accompanied by algorithmic pip
elines which automate the process of learning rich
representations from the data. To achieve princip
led regularisation it is essential to communicate
the uncertainty across the different stages of the
pipelines and the different components of the gra
phical models. Therefore\, the talk will also pres
ent a set of mathematical developments which achie
ve this through variational inference. The above c
oncepts and algorithms will be demonstrated with e
xamples from computer vision (high-dimensional vid
eo\, images)\, robotics (motion capture data\, hum
anoid robotics) and dynamical systems.
LOCATION:Engineering Department\, CBL Room BE-438
CONTACT:
END:VEVENT
END:VCALENDAR