Reviews: Dual Decomposed Learning with Factorwise Oracle for Structural SVM of Large Output Domain

This paper introduces an approach to learning of structural SVMs based on a direction method of multipliers. The starting point is to bring the primal form of the learning objective into a dual-decomposed representation (eq. This representation is frequently used for inference in intractable graphical models, and has meanwhile also been exploited successfully for structured learning (e.g. in [13,14]). Basically, the problem is broken down into maximization of individual factors, as well as minimization with respect to the model parameters as well as Lagrange variables coupling the factors. Approaches then differ in how the coupling constraint are handled, and in how the minimization with respect to the model parameters is conducted.