Bundle method solver for structured output learning
by Michal Uřičář for Shogun Machine Learning Toolbox
Learning of the structured output classifiers leads to solving a convex minimization problem which is not tractable by standard algorithms. A significant effort in ML community has been put to development of specialized solvers among which the Bundle Method for Risk Minimization (BMRM), implemented e.g. in popular StructSVM, is the current the state-of-the-art. The BMRM is a simplified variant of bundle methods which are standard tools for non-smooth optimization. The simplicity of the BMRM is compensated by its reduced efficiency. Experiments show that a careful implementation of the classical bundle methods perform significantly faster (speedup ~ 5-10) than their variants (like BMRM) adopted by ML community. The goal will be an OS library implementing the classical bundle method for the SO learning and its integration to Shogun.