BEGIN:VCALENDAR
VERSION:2.0
PRODID:icalendar-ruby
CALSCALE:GREGORIAN
METHOD:PUBLISH
BEGIN:VTIMEZONE
TZID:Europe/Vienna
BEGIN:DAYLIGHT
DTSTART:20190331T030000
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:CEST
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20181028T020000
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
TZNAME:CET
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20191210T171947Z
UID:5c265de4c2cce904949525@ist.ac.at
DTSTART:20190128T100000
DTEND:20190128T110000
DESCRIPTION:Speaker: Laurence Aitchison\nhosted by Peter Jonas\nAbstract: H
ere\, I present three projects using principled\, Bayesian treatments of u
ncertainty to address pressing problems in modern deep learning.1.) Bayesi
an inference in modern convolutional neural networks.Bayesian inference ov
er the parameters of a neural networks is critical to improve generalisati
on performance\, and to reason about what the network does not know due to
limited training data. However\, Bayesian inference in typical neural net
works is impossible (at least without severe approximations) due to the sh
eer number of parameters in these networks. Here\, we show exact inference
is possible in state-of-the-art convolutional networks\, if we take the l
imit of infinitely many convolutional filters (at which point the outputs
follow a Gaussian process). The network obtains 0.84 % classification erro
r on MNIST\, a new record for a Gaussian Process method.2.) Neural network
optimization as Bayesian inferenceNeural network optimization methods fal
l into two broad classes: adaptive methods such as RMSprop and non-adaptiv
e methods such as stochastic gradient descent (SGD). This presents a probl
em for practitioners: which method should use on a particular problem? Or
even should you use an adaptive method on some parameters\, and a non-adap
tive method on others? Here\, we resolve this issue\, by deriving a Bayesi
an gradient descent rule that adaptively transitions between adaptive and
non-adaptive behaviour. This method provides insight into when we might ex
pect adaptive and non-adaptive methods to be most useful\, and is superior
to standard neural network optimization methods in practice.3.) Bayesian
inference in deep graphical models Graphical models are a powerful languag
e to encode our knowledge of the dependency (or even causal) structure in
data. However\, graphical modelling has fallen out of favour recently\, du
e to the success of often unstructured deep-learning. Here\, we show that
it is possible to combine deep learning and graphical models to form "deep
graphical models". To perform inference\, we combine strategies from deep
learning (variational autoencoder recognition models) with strategies fro
m graphical modelling (message passing). The resulting inference schemes g
ive considerably improved performance over a vanilla deep-learning inferen
ce strategy.
LOCATION:Mondi Seminar Room 2\, Central Building\, IST Austria
ORGANIZER:tguggenb@ist.ac.at
SUMMARY:Bayesian inference and deep learning
URL:https://talks-calendar.app.ist.ac.at/events/1719
END:VEVENT
END:VCALENDAR