Neural networks: from statistical physics to neurophysiology.

Summary Several works concerning neural networks are presented. In the first part, statistical physics techniques are applied to several situations. One of them is the study of categorization and the occurrence of prosopagnosia in neural networks with structured attractors. In the second part, more realistic neural networks are studied. First, the conditions on the input stimuli are determined so that they can be learned, and then a model of an autonomously learning network is studied. This network includes analog neurons and a stochastic learning dynamic on synaptic efficiencies. Finally, the same type of network, studied in the context of learning a fixed sequence of stimuli, allows us to confront the model in a fruitful way with the neurophysiology experiments of miyashita.