A general framework for generalization : the double-minimization method
Un cadre général pour la généralisation : la méthode de double simplification
Résumé
The paper proposes a general framework for generalization : the research of an excellent network can be seen as the research of a space in a lattice of functional spaces (structures of networks) and an execellent function in this space (the coefficients or degrees of freedom). In order to illustrate this framework, we use it in the case of orthonormal basis nets. Our experimentation leads to a crucial constatation : the evidence that a good capacity of generalization supposes a good capacity to change the structure of the network.
On montre qu'une méthode de double simplification peut améliorer la capacité de généralisation d'un algorithme d'apprentissage.