loopy learning

Brendan Frey (frey@dendrite.uwaterloo.ca)
Wed, 9 Jun 1999 08:39:48 -0400

It turns out that loopy propagation, or "turboinference", in Gaussian
networks can be used effectively for INFERENCE and LEARNING:

http://www.cs.toronto.edu/~frey/papers/tfa-nc99.abs.html

In this paper (submitted for publication in May), I show that
iterative probability propagation in factor analysis networks has a fixed
point and I give an eigenvalue condition for global convergence.

I also show that iterative propagation can be used for learning factor
analyzer networks and give results on

face recognition.

Brendan.