Learning Randomly Perturbed Structured Predictors
                 for Direct Loss Minimization

                     Hedda Cohen Indelman 1 Tamir Hazan 1

             Abstract                gradients of two predicted structures: over the original loss
    Direct loss minimization is a popular approach for     function and over a perturbed loss function. This approach
    learning predictors over structured label spaces.     implicitly uses the data distribution to smooth the loss func-
    Th ...