Navigating the Trade-Off between Multi-Task Learning and Learning to Multitask in Deep Neural Networks