Learning in ultrametric committee machines (Q1938814)

USMs are understood as a class of feed-forward, fully connected multilayer neural networks with binary units, ultrametric matrix of overlaps and with a processing of inputs according to some specific rule. The learning-by-examples in such a system is investigated and this is realized by means of statistical mechanics techniques and interpretations. The function expressing the free energy of the system, the saddle point equation and the generalization error are computed. There is a discussion on their behavior in different regimes characterized by the number of examples and complexity of the neural network.