– Yann Guermeur (LORIA/Lorraine Université) : Guaranteed risk for large margin multi-category classifiers

Carte non disponible

Date(s) - 12 juin 2015

Catégories Pas de Catégories

Title : Guaranteed risk for large margin multi-category classifiers\n \nAbstract : In the framework of agnostic learning, the two basic parameters of a multi-class discrimination problem are the sample size m and the number of categories C. In 2007, we contributed to the Vapnik-Chervonenkis theory of large margin multi-category classifiers by introducing the appropriate class of generalized Vapnik-Chervonenkis dimensions : the class of gamma-psi-dimensions. The guaranteed risk we derived exhibited a suboptimal ln(m) / m^1/2 convergence rate. In 2012, Mohri and his co-authors obtained the optimal 1/m^1/2 rate, with a control term growing quadratically with C. In this talk, we prove that this result can be improved upon, by establishing a bound of equal convergence rate and better dependency on C, namely a O(C^3/2).[