Output details

Learning in ultrametric committee machines

<24> This paper aims at the goal of understanding the computational capabilities of feed-forward neural networks beyond the perceptron. It shows that full generalization is only achieved when the student network has the same architecture and capabilities as the teacher network. All the learning processes observed present a transition separating a non-specialized and a specialized phase. Here specialization refers to the values of the student's synaptic weights and their similarity with the corresponding weights in the teacher. This work was the result of an ongoing collaboration with Prof. Leonardo Franco, Universidad de Malaga, Spain.