Reviews: Learning to Multitask

This paper presents an end-to-end framework for multitask learning called L2MT. The L2MT is trained on previous multitask problems, and selects the best multitask model for a new multitask problem. Authors also proposed a effective algorithm for selecting the best model on test data. Pros: - This is an clean and straightforward framework for multitask learning, and it is end-to-end. Cons: - The'label' in this learning task is the relative test error of eps_{MTL}/eps_{STL}. In my understanding, eps_{STL} is roughly measuring the level of difficulty of a multitask problem, but is there a better baseline to use here?