Faster Training of Structural SVMs with Diverse M-Best Cutting-Planes

Training of Structural SVMs involves solving a large Quadratic Program (QP). One popular method for solving this optimization problem is a cutting-plane approach, where the most violated constraint is iteratively added to a working-set of constraints. Unfortunately, training models with a large number of parameters remains a time consuming process. This paper shows that significant computational savings can be achieved by generating and adding multiple highly violated constraints at every iteration of the training algorithm. We show that generation of such diverse M-Best cutting-planes involves extracting diverse M-Best solutions from the loss-augmented score function. Our experiments on image segmentation and protein side-chain prediction show that the proposed approach can lead to significant computational savings, e.g., > 60% reduction in the number of training iterations. Finally, our results suggest that even greater savings are possible for more expressive models involving a larger number of parameters.