[en] Several learning algorithms in classification and structured prediction are formulated as large scale optimization problems. We show that a generic iterative reformulation and resolving strategy based on the progressive hedging algorithm from stochastic programming results in a highly parallel algorithm when applied to the large margin classification problem with nonlinear kernels. We also underline promising aspects of the available analysis of progressive hedging strategies.
Systems and Modeling Research Unit
DYSCO (Dynamical Systems, Control, and Optimization); PASCAL2 Network of Excellence