New release of Bayesian NN software

[Radford Neal]

                        New Release of Software for

                    BAYESIAN LEARNING FOR NEURAL NETWORKS

                     Radford Neal, University of Toronto

 A new version of my software for Bayesian learning of models based on
 multilayer perceptron networks, using Markov chain Monte Carlo methods, 
 is now available on the Internet.  This software implements the methods 
 described in my Ph. D. thesis, "Bayesian Learning for Neural Networks",
 which is now available from Springer-Verlag (ISBN 0-387-94724-8).  Use 
 of the software is free for research and educational purposes.

 The software supports models for regression and classification problems
 based on networks with any number of hidden layers, using a wide variety
 of prior distributions for network parameters and hyperparameters.  The
 advantages of Bayesian learning include the automatic determination of 
 "regularization" parameters, without the need for a validation set, 
 avoidance of overfitting when using large networks, and quantification of 
 the uncertainty in predictions.  The software implements the Automatic 
 Relevance Determination (ARD) approach to handling inputs that may turn 
 out to be irrelevant (developed with David MacKay).  For problems and 
 networks of moderate size (eg, 200 training cases, 10 inputs, 20 hidden 
 units), full training (to the point where one can be reasonably sure that 
 the correct Bayesian answer has been found) typically takes several hours 
 to a day on our SGI machine.  However, quite good results, competitive 
 with other methods, are often obtained after training for under an hour.  
 (Of course, your machine may not be as fast as ours!)

 The software is written in ANSI C, and has been tested on SGI and Sun
 machines.  Full source code is included.  This new release is not 
 radically different from the release of a year ago, but it does contain
 a number of enhancements to both the programs and the documentation, so 
 it is probably worth your while to upgrade if you are using the old
 version.  The new version is not quite upwardly compatible with the old,
 but converting old scripts should be very easy.
 
 You can obtain the software via my home page, at URL

             http://www.cs.toronto.edu/~radford/

 If you have any problems obtaining the software, please contact me at one 
 of the addresses below.

 ---------------------------------------------------------------------------
 Radford M. Neal                                      radford at cs.toronto.edu
 Dept. of Statistics and Dept. of Computer Science  radford at stat.toronto.edu
 University of Toronto                    http://www.cs.toronto.edu/~radford
 ---------------------------------------------------------------------------