UC San Diego

Boosting algorithms produce accurate predictors for complex phenomena by welding together collections of simple predictors. In the classical method AdaBoost, as well as its immediate variants, the welding points are determined by convex optimization; unlike typical applications of convex optimization in machine learning, however, the AdaBoost scheme eschews the usual regularization and constraints used to control numerical and statistical properties. On the other hand, the data and simple predictors impose rigid structure on the behavior of AdaBoost variants, and moreover convex duality provides a lens to resolve this rigidity. This structure is fundamental to the properties of these methods, and in particular leads to numerical and statistical convergence rates