NeurIPS 2020

Differentiable Top-k with Optimal Transport

Meta Review

The reviews for this paper were positive overall. The paper presents a smoothing technique for the top-k operator via the addition of an entropic regularization. The proposed smoothed counterpart departs from previous proposals in that it may be continuous yet non-smooth. arg-max. The contributions were considered 'well-motivated' and 'highly relevant'. The ease of use of the PyTorch implementation uploaded along with the paper was particularly appreciated. We recommend to take the reviewers' comments and suggestions into account while preparing the camera ready final version of the paper. The authors may pay particular attention to delineate clearly the contributions of the paper from previous papers on the subject (as well as the originality of their contributions). Accept.