Abstract
The approach of the previous chapter is extended to the derivation of the model evidence as a measure for model weighting and selection. The idea is to integrate out the hyperparameters in the likelihood term by Gaussian approximation, which requires the derivation of the Hessian at the mode. The resulting expression for the model evidence is found to be an intuitively plausible generalisation of the results obtained by MacKay for Gaussian homoscedastic noise on the target. The nature of the various Ockham factors included in the evidence is discussed. The chapter concludes with a critical evaluation of the numerical inaccuracies inherent in this scheme.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Preview
Unable to display preview. Download preview PDF.
Reference
The given derivation is only intuitive reasoning; a proper mathematical treatment of this subject can be found in [9], Chapter 1. In fact, the prior of equation 2 (11.29) is not uninformative in the output weights a, the implications of which are discussed on page 176.
Strictly speaking, this equation only holds locally within intervals of length , In , and In . However, these details have no influence on the following analysis and will therefore not be further considered here.
Note that W (the number of weights) is denoted by m in MacKay’s work.
See also [7], chapter 10.
Author information
Authors and Affiliations
Rights and permissions
Copyright information
© 1999 Springer-Verlag London Limited
About this chapter
Cite this chapter
Husmeier, D. (1999). The Bayesian Evidence Scheme for Model Selection. In: Neural Networks for Conditional Probability Estimation. Perspectives in Neural Computing. Springer, London. https://doi.org/10.1007/978-1-4471-0847-4_11
Download citation
DOI: https://doi.org/10.1007/978-1-4471-0847-4_11
Publisher Name: Springer, London
Print ISBN: 978-1-85233-095-8
Online ISBN: 978-1-4471-0847-4
eBook Packages: Springer Book Archive