NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:1842

This paper addresses a method for learning meta-curvature information from various similar tasks, which is claimed to help better generalization and fast model adaptation. A tensor decomposition is applied to the meta-curvature to scale up the method. All of reviewers agree that the paper is well written and the idea on the tensor decomposition of the curvature is interesting. However most of reviewers did not pinpoint a few issues. First of all, it is not clear why the method provides better generalization. Second, the meta-curvature in this paper is similar to the idea of he MT-net [Lee and Choi, 2018], while the novelty here is the application of tensor decomposition for scaling up. During the discussion period, unfortunately none of reviewers did express further feedback. There was no "strong support" or "negative support" either. Anyway, the paper is slightly above the threshold, deserved to be presented at NeurIPS this year.