Structure and Contributions of the Thesis

This thesis is organised as follows. Chapter  discusses previous work on extensions of factor analysis that can be used for unsupervised learning. Chapter  gives an overview of Bayesian ensemble learning which is the essential theoretical background. Building blocks and their usage with ensemble learning are described in Chapter . The model structure used in hierarchical nonlinear factor analysis with variance modelling (HNFA+VM) is built from these blocks in Chapter . Chapter  describes the algorithm that is used to let the model learn from the data.

Two sets of experiments were conducted using HNFA+VM. In Chapter , an artificial bars problem is analysed and in Chapter , the model is applied on natural image data. Finally, the benefits, restrictions and applications of the model and future work are discussed in Chapter .

The update rule for the Gaussian variable with nonlinearity has been developed by the author. The experiments as well as the code for preprocessing and parts of the learning procedure like initialisation, pruning, regeneration and rebooting, have beed developed by the author.