How would you design a loss function for a multi-task learning problem where the tasks have different scales?
