Artificial neural networks

When regarding (artificial) neural networks as biologically inspired structures showing several of the brains' characteristics, unsupervised learning (i.e. learning without an external guidance) gains major attention as it is known for its biological plausibility (as opposed to supervised learning involving a teacher).

Unsupervised learning neural networks aim at building useful representations from the external data world by learning from input examples (regarded as sensory inputs in biological terminology). These representations may be used for probabilistic reasoning, solving classification tasks, modeling theories of human perception, etc. Representations considered as useful are typically those that form topographic maps, model the input data density, construct dimensionality reducing mappings, identify clusters - depending on what purpose the unsupervised learning network serves. We are interested in studying the links between representations and the learning target focussing on the computational aspects.

In particular, we look into graphical models forming a successful probabilistic modeling approach encoding relationships among a set of random variables and provide a representation for the joint probability distribution over these variables. The advantages of the graphical formalism have their origins in probability theory and graph theory, the structural modularity favoring parallel computations, and its clearness.

Addressing the model selection problem of unsupervised trained structures by using evolutionary algorithms suggests itself, as we remain in the area of parallel computing inspired by nature.