Which Tasks Should Be Learned Together in Multi-task Learning?

March 2020

tl;dr: Answers this question: which tasks are better trained with others?

Overall impression

The paper uses the dataset from CVPR 2018 best paper taskonomy, which studies the task transferability. Task grouping studies the multi-task learnability. The paper founds that they correlate inversely.

The goal of multitask learning is two-fold.

One of the key insights from the paper is:

The inclusion of an additional task in a network can potentially improve the accuracy that can be achieved on the existing tasks, even though the performance of the added task might be poor.

Key ideas

Technical details