Duchi and Singer [24] proposed a boosting method for multi-class classification problems by utilizing the structural sparsity of model parameters. They claimed that the method can be generalized for multi-task learning.