Section: New Results

Structured Learning and Inference

Paticipants: Jiaqian Yu, Matthew Blaschko

We have developed computationally efficient structured output prediction methods for learning with non-modular losses [19] , [29] , [40] . We both demonstrate the feasibility of learning with submodular losses, as well as show that learning with multiple correct outputs can lead to NP-hard optimization problems even when learning with a single correct output is feasible.