NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:2222
Title:SCAN: A Scalable Neural Networks Framework Towards Compact and Efficient Models


		
Reviewers had extensive discussions on this paper during the discussion phase. One comment was that the paper appeared to justify exclusively the empirical performance over methodological contributions. Other reviewers felt that the significant improvement in performance warranted publication. I concur on this latter view. As a note, and as reviewers have pointed out as well, there are many related works that the authors should contrast in their revised version. Some of these include: Adaptive Neural Networks for Efficient Inference, FRACTALNET, Feedback Networks, and NestedNet. Clearly, the early exit (or early prediction, drop path) is totally not a new idea. While those papers appear to have similar ideas, the performance achieved in the submitted paper is significant.