Ensemble Learning and Evidence Maximization

Ensemble learning by variational free energy minimization is a tool introduced to neural networks by Hinton and van Camp in which learning is described in terms of the optimization of an ensemble of parameter vectors. The optimized ensemble is an approximation to the posterior probability distribution of the parameters. This tool has now been applied to a variety of statistical inference problems. In this paper I study a linear regression model with both parameters and hyper-parameters. I demonstrate that the evidence approximation for the optimization of regularization constants can be derived in detail from a free energy minimization viewpoint. A new tool has recently been introduced into the eld of neural networks and statistical inference. In traditional approaches to neural networks, a single parameter vector w is optimized by maximum likelihood or penalized maximum likelihood. In the Bayesian interpretation , these optimized parameters are viewed as deening the mode of a posterior probability distribution P (wjD; H) (given data D and model assumptions H), which can be approximated, with a Gaussian distribution ~ P for example (MacKay 1992b), in order to obtain predictive distributions and optimize model control parameters. The new concept introduced by Hinton and van Camp (1993) is to work in terms of an approximating ensemble Q(w;), that is, a probability distribution over the parameters, and optimize the ensemble (by varying its own parameters) so that it approximates the posterior distribution of the parameters P (wjD; H) well. The objective function chosen to measure the quality of the approximation is a variational free energy, 1