| Literature DB >> 16764513 |
Geoffrey E Hinton1, Simon Osindero, Yee-Whye Teh.
Abstract
We show how to use "complementary priors" to eliminate the explaining-away effects that make inference difficult in densely connected belief nets that have many hidden layers. Using complementary priors, we derive a fast, greedy algorithm that can learn deep, directed belief networks one layer at a time, provided the top two layers form an undirected associative memory. The fast, greedy algorithm is used to initialize a slower learning procedure that fine-tunes the weights using a contrastive version of the wake-sleep algorithm. After fine-tuning, a network with three hidden layers forms a very good generative model of the joint distribution of handwritten digit images and their labels. This generative model gives better digit classification than the best discriminative learning algorithms. The low-dimensional manifolds on which the digits lie are modeled by long ravines in the free-energy landscape of the top-level associative memory, and it is easy to explore these ravines by using the directed connections to display what the associative memory has in mind.Entities:
Mesh:
Year: 2006 PMID: 16764513 DOI: 10.1162/neco.2006.18.7.1527
Source DB: PubMed Journal: Neural Comput ISSN: 0899-7667 Impact factor: 2.026