multi-task learning (MTL) is to learn one single model that performs multiple
tasks for achieving good performance on all tasks and lower cost on
computation. Learning such a model requires to jointly optimize losses of a set
of tasks with different difficulty levels, magnitudes, and c