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:Microsoft Research Cambridge\, public talks
SUMMARY:How to force unsupervised neural networks to disco
 ver the right representation of images - Geoffrey 
 Hinton\, University of Toronto
DTSTART;TZID=Europe/London:20110623T110000
DTEND;TZID=Europe/London:20110623T120000
UID:TALK31612AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/31612
DESCRIPTION:One appealing way to design an object recognition 
 system is to define objects recursively in terms o
 f their parts and the required spatial relationshi
 ps between the parts and the whole. These relation
 ships can be represented by the coordinate transfo
 rmation between an intrinsic frame of reference em
 bedded in the part and an intrinsic frame embedded
  in the whole. This transformation is unaffected b
 y the viewpoint so this form of knowledge about th
 e shape of an object is viewpoint invariant. A nat
 ural way for a neural network to implement this kn
 owledge is by using a matrix of weights to represe
 nt each part-whole relationship and a vector of ne
 ural activities to represent the pose of each part
  or whole relative to the viewer. The pose of the 
 whole can then be predicted from the poses of the 
 parts and\, if the predictions agree\, the whole i
 s present. This leads to neural networks that can 
 recognize objects over a wide range of viewpoints 
 using neural activities that are ``equivariant'' r
 ather than invariant: as the viewpoint varies the 
 neural activities all vary even though the knowled
 ge is viewpoint-invariant. The ``capsules'' that i
 mplement the lowest-level parts in the shape hiera
 rchy need to extract explicit pose parameters from
  pixel intensities and these pose parameters need 
 to have the right form to allow coordinate transfo
 rmations to be implemented by matrix multiplies. T
 hese capsules are quite easy to learn from pairs o
 f transformed images if the neural net has direct\
 , non-visual access to the transformations\, as it
  would if it controlled them. (Joint work with Sid
 a Wang and Alex Krizhevsky)
LOCATION:Small lecture theatre\, Microsoft Research Ltd\, 7
  J J Thomson Avenue (Off Madingley Road)\, Cambrid
 ge
CONTACT:Microsoft Research Cambridge Talks Admins
END:VEVENT
END:VCALENDAR