|
NeuroCOLT
Technical Report NC-TR-96-012
Using Bayesian
Methods for Avoiding Overfitting and for Ranking Networks in Multilayer
Perceptrons Learning
Michel
de Bollivier
EC Joint Research Centre
Italy
Domenico
Perrotta
EC Joint Research Centre and Ecole Normale Superieure de Lyon
France
Abstract
This work is an experimental attempt to determine whether the Bayesian
paradigm could improve Multi-Layer Perceptrons (MLPs) learning methods.
In particular, we exper iment here the paradigm developed by D. MacKay
(1992). The paper points out the main or critical points of MacKay's
work and introduces very practical points of Bayesian MLPs, having
in mind future applications. Then, Bayesian MLPs are used on three
public classification databases and compar ed to other methods.
Download Compressed
Postscript
|