Reviews: Consistent Multitask Learning with Nonlinear Output Relations

The paper tackles multi-task learning problems where there are non-linear relationships between tasks. The relationships between tasks is encoded as a set of non-linear constraints that the outputs of each task must satisfy (e.g . In a nutshell, he proposed technique can be summarized as: use kernel regression to make predictions for each task independently, then project the prediction vector onto the constrained set. Overall, I like the idea of being able to take advantage of non-linear relationships between tasks. However, I am not sure how practical it is to specify the non-linear constraints between tasks in practice.