BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Learning visual representations - Dr. Andrea Vedaldi\, University 
 of Oxford
DTSTART:20150310T140000Z
DTEND:20150310T150000Z
UID:TALK58309@talks.cam.ac.uk
CONTACT:Rachel Fogg
DESCRIPTION:Learnable representations\, and deep convolutional neural netw
 orks (CNNs) in particular\, have become the preferred way of extracting vi
 sual features for image understanding tasks\, from object recognition and 
 detection to semantic segmentation. \n\nIn this talk I will discuss severa
 l recent advances in deep representations for computer vision. First\, I w
 ill review modern CNN architectures and their training. Then\, I will illu
 strate state-of-the-art networks using an example in text spotting. In thi
 s example I will show that\, by using only synthetic data and a sufficient
 ly large deep model\, it is possible to learn to directly map image region
 s to entire words\, effectively training 90k image classifiers that achiev
 e state-of-the-art text spotting performance. I will also briefly touch on
  other applications of deep learning in object recognition and discuss fea
 ture universality and transfer learning.\n\nIn the last part of the talk I
  will move to the problem of understanding deep networks\, which remain\, 
 by and large\, black boxes\, presenting two recent results. The first one 
 are visualisation techniques to investigate which visual information is re
 tained or learned by visual representations. The second one is a method th
 at allows investigating how geometric transformations are represented in a
  CNN\, as well as establishing whether two CNNs\, learned on different tas
 ks\, are in fact equivalent.\n
LOCATION: Cambridge University Engineering Department\, Lecture Room 3B
END:VEVENT
END:VCALENDAR