Wake-sleep algorithm

From Wikipedia, the free encyclopedia
Jump to: navigation, search

The wake-sleep algorithm is an unsupervised learning algorithm for a multilayer neural network. Training is divided into two phases, "wake" and "sleep". In the "wake" phase, neurons are driven by recognition connections (connections from what would normally be considered an input to what is normally considered an output), while generative connections (those from outputs to inputs) are modified to increase the probability that they would reconstruct the correct activity in the layer below (closer to the sensory input). In the "sleep" phase the process is reversed: neurons are driven by generative connections, while recognition connections are modified to increase the probability that they would produce the correct activity in the layer above (further from sensory input).

See also[edit]

External links[edit]