Efficient algorithms for max-margin structured classification
We present a general and efficient optimisation methodology for for max-margin sructured classification tasks. The efficiency of the method relies on the interplay of several techiques: marginalization of the dual of the structured SVM, or max-margin Markov problem; partial decomposition via a gradient formulation; and finally tight coupling of a max-likelihood inference algorithm into the optimization algorithm, as opposed to using inference as a working set maintenance mechanism only. The tight coupling also allows fast approximate inference to be used effectively in the learning. The generality of the method follows from the fact that changing the output structure in essence only changes the inference algorithm, that is, the method can almost be used in 'plug and play' fashion.