BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:A Fast Learning Algorithm for Deep Belief Nets - Geoffrey E. Hinto
 n\, University of Toronto
DTSTART:20050615T140000Z
DTEND:20050615T150000Z
UID:TALK4374@talks.cam.ac.uk
CONTACT:Phil Cowans
DESCRIPTION:I will show how ``complementary priors'' can be used to elimin
 ate the\nexplaining away effects that make inference difficult in\ndensely
 -connected belief nets that have many hidden layers.  Using\ncomplementary
  priors\, I will derive a fast\, greedy algorithm that can\nlearn deep\, d
 irected belief networks one layer at a time\, provided the\ntop two layers
  form an undirected associative memory. The fast\, greedy\nalgorithm is us
 ed to initialize a slower learning procedure that\nfine-tunes the weights 
 using a contrastive version of the wake-sleep\nalgorithm. After fine-tunin
 g\, a network with three hidden layers forms\na very good generative model
  of the joint distribution of handwritten\ndigit images and their labels. 
 This generative model gives better\nclassification performance than discri
 minative learning\nalgorithms. The low-dimensional manifolds on which the 
 digits lie are\nmodeled by long ravines in the free-energy landscape of th
 e top-level\nassociative memory and it is easy to explore these ravines by
  using\nthe directed connections to display what the associative memory ha
 s in\nmind.\n(Joint work with Simon Osindero and Yee-Whye Teh)
LOCATION:Ryle Seminar Room\, Cavendish Laboratory
END:VEVENT
END:VCALENDAR