A BP-Based Algorithm for Performing Bayesian
Inference in Large Networks of Perceptron-Type
Lecture within the tutorial Statistical Methods in Learning
Speaker: Yoshiyuki Kabashima (Tokyo Institute of Technology)
The Bayesian approach provides the optimal performance
in many inference problems. However, the necessary cost for computation
sometimes prevents it from being practical. In my talk, it is shown
that one can develop a practical algorithm to approximately perform
the Bayesian inference in large perceptron-type networks by
introducing methods and notions of statistical mechanics to
Pearl's belief propagation (BP) although the direct application
of BP to such problems is computationally difficult.
An application of the developed algorithm to a problem that
arises in a wireless communication system is also mentioned.
This is joint work with Shinsuke Uda.
©Copyright 2004 Author