BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Machine Learning Meetup  - Marc Brockschmidt (MSR Cambridge) and V
 incent Dutordoir (Prowler)
DTSTART:20200131T173000Z
DTEND:20200131T193000Z
UID:TALK139006@talks.cam.ac.uk
CONTACT:Microsoft Research Cambridge Talks Admins
DESCRIPTION:The aim of this meetup in Microsoft Research\, Cambridge is to
  bring together people interested in machine learning and artificial intel
 ligence\, be it applied or theory.  The meetup features two 15-minute rese
 arch talks followed by a small research poster session during which we ser
 ve light refreshments. \n\nThis Friday\, we will have two exciting talks:\
 n\nSpeaker: Marc Brockschmidt (MSR Cambridge)\n\nTitle: Editing Sequences 
 by Copying Spans\n\nAbstract: Neural sequence-to-sequence models are findi
 ng increasing use in editing of documents\, for example in correcting a te
 xt document or repairing source code. In this paper\, we argue that existi
 ng seq2seq models (with a facility to copy single tokens) are not a natura
 l fit for such tasks\, as they have to explicitly copy each unchanged toke
 n. We present an extension of seq2seq models capable of copying entire spa
 ns of the input to the output in one step\, greatly reducing the number of
  decisions required during inference. This extension means that there are 
 now many ways of generating the same output\, which we handle by deriving 
 a new objective for training and a variation of beam search for inference 
 that explicitly handle this problem.\n\nSpeaker: Vincent Dutordoir (Prowle
 r)\n\nTitle: Bayesian Image Classification with Deep Convolutional Gaussia
 n Processes\n\nAbstract: There is a lot of focus on Bayesian deep learning
  at the moment\, with many researchers tackling this problem by building o
 n top of neural networks and making the inference look more Bayesian. In t
 his talk\, I'm going to follow a different strategy and use a Gaussian pro
 cess\, which is a well-understood probabilistic method with many attractiv
 e properties\, as a primitive building block to construct fully Bayesian d
 eep learning models. We show that the accuracy of these Bayesian methods\,
  and the quality of their posterior uncertainties\, depend strongly on the
  suitability of the modelling assumptions made in the prior\, and that Bay
 esian inference by itself is often not enough. This motivates the developm
 ent of a novel convolutional kernel\, which leads to improved uncertainty 
 and accuracy on a range of different problems.\n\n
LOCATION:Auditorium\, Microsoft Research Ltd\, 21 Station Road\, Cambridge
 \, CB1 2FB
END:VEVENT
END:VCALENDAR