next up previous contents
Next: Model Up: Bayesian Ensemble Learning for Previous: Algorithms


The motivation for developing methods tailored for unsupervised learning is that to learn large hierarchical models, unsupervised learning is the most promising approach. A characteristic property of unsupervised learning is the potentially large amount of unknown variables. This is because the latent variables need to be estimated for each observation separately. In Bayesian learning, the posterior distribution of these unknown variables requires to be estimated.

In this section, a nonlinear extension of the factor analysis model is used as an example of how ensemble learning can be applied to unsupervised learning of this kind of model and why it should be used in the first place. Ensemble learning is discussed at length in publication IV and a detailed description of its application to nonlinear factor analysis by MLP network is given in publication V. In this section, only the main points are summarised.


Harri Valpola