Statistical physics of neural networks Original Research Article

The cooperative behaviour of interacting neurons and synapses is studied using models and methods from statistical physics. The competition between training error and entropy can lead to discontinuous properties of neural networks. This is demonstrated for a few examples: Attractor networks, storage by synaptic dilution, learning from examples, generalization of multilayer networks, learning without generalization, specialization, Bayesian estimate of the structure of high dimensional data.