The Helmholtz machine is a new unsupervised learning architecture that
uses top-down connections to build probability density models of inpu
t and bottom-up connections to build inverses to those models. The wak
e-sleep learning algorithm for the machine involves just the purely lo
cal delta rule. This paper suggests a number of different varieties of
Helmholtz machines, each with its own strengths and weaknesses. and r
elates them to cortical information processing. Copyright (C) 1996 Els
evier Science Ltd.