Two different experimental settings are used in this thesis. First an extension to the artificial bars problem [12] is studied to demonstrate that the algorithm actually finds a model that is similar to the one used to generate the data. Small image patches are generated by randomly inserting horisontal and vertical bars and areas with increased noise level to an image.
The second experiment setting consists of analysing patches of natural gray-scale images. In this case there is no right answer, but the resulting model can be compared to what biologists know about the human brain. Hyvärinen and Hoyer [30,29,31] applied independent component analysis (ICA) and its extensions to natural images. The basic ICA leads to emergence of simple cell properties and the extensions lead to emergence of both topography and invariances similar to complex cell properties.
Frey and Jojic estimated a mixture model for images with a fixed set of transformations [16]. Using 100 frames of head-and-shoulder video sequences of a person walking across a higly cluttered background and a fixed set of translational invariances, they could capture clusters that presented the person with different poses, while background was interpreted as noise.