An Efficient Method for Large Margin Parameter Optimization in Structured Prediction Problems
Huizhen Yu and Juho Rousu
Department of Computer Science.
We consider structured prediction problems with a parametrized linear prediction function, and the associated parameter optimization problems in large margin type of discriminative
training. We propose a dual optimization approach which uses the restricted simplicial decomposition method to optimize a reparametrized dual problem. Our reparametrization reduces the
dimension of the space of the dual function to one that is linear in the number of parameters and training examples, and hence independent of the dimensionality of the prediction outputs.
This in conjunction with simplicial decomposition makes our approach efficient. We discuss the connections of our approach with related earlier works, and we show its advantages.