NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:941
Title:Self-Supervised Generalisation with Meta Auxiliary Learning


		
This paper addresses a method to improve a primary task by jointly learning auxiliary tasks. A unique contribution in this paper is in the label generation network which is trained in a self-supervised manner to generate labels for the auxiliary task. A way to jointly train both multi-task network and label-generation network is similar to MAML. In that sense, the method is referred to as meta auxiliary learning, but such name could be misleading. All of reviewers agree that the paper has an interesting idea. On the other hand, there are concerns on experiments where only marginal improvement over single task learning is shown. During the discussion, one reviewer raised his/her score, but they still have concerns that improvements are marginal in image classification tasks.