NeurIPS 2020

Fast Fourier Convolution


Meta Review

Reviewers opinion was split on this paper. I am slightly less concerned with issues of novelty than some reviewers are - compositional methods (things that combine existing approaches) are often criticised for novelty but the approaches are commonly non-obvious a priori, and in this field almost all things are compositional. I think it is sufficiently non-obvious that the combination in this paper would provide the performance benefit it does. Hence it is of interest to practitioners in the field. The authors should temper their claims a bit about what they are the first to do, as mentioned by a couple of reviewers, and the authors must find a way to include their ablation tests in the rebuttal into the main paper. I appreciate the work done in the rebuttal to address reviewers' concerns. In summary, I felt the reviewers who gave negative scores also made important positive points and that the reasons given for low scores were either countered in the rebuttal, or not sufficiently important to prevent publication here. No doubt some reviewers will disagree, but there is always going to be disagreement in academic discourse!