A flexible scheme to learn to classify, to infer and to predict things simply by observing examples, would be invaluable. Visualisation techniques help humans to perceive large data sets at a glance, but there is still trouble in perceiving data with high intrinsic dimensionality. An automatic system is not limited to any strict number of dimensions. As the processing power and the amount of available raw data grows, the greatest cost of analysation is the amount of human intervention.
Analysing images is an effective way to study these models since the results are relatively easy to interpret. Images can be considered as high dimensional data by thinking that each pixel is a component of a data vector. One observation vector is thus an image patch. The human brain is specialised among many other things to interpreting visual observations. Fairly sophisticated computer vision methods for certain tasks like for face recognition [68] exist, but they are highly specialised and thus incapable of adapting to new situations. The capabilities of such a system are limited by the engineering work, which easily builds up to the system until it is an incomprehensible mass of finely tuned rules.