Statistical physics of learning:Phase transitions in multilayered neural networks

The statistical physics of disordered systems provides tools for the investigation of learning processes in adaptive information processing. The methods and objectives of this approach are exemplified in terms of a specific model scenario: the supervised learning of a rule with a multilayered neural network. The model exhibits a discontinuous dependence of the student performance on the number of example data. This phenomenon can be interpreted as a symmetry breaking phase transition, which results from the competition of (formal) energy and entropy