NeurIPS 2020

On Testing of Samplers

Meta Review

This paper introduces a method for testing samplers with discrete distributions as targets. The authors demonstrate its usefulness in theory and experiments. The reviewers seem to be in agreement that the main strength of the paper is its novelty in addressing an important problem (testing of samplers). While the presented method does not (as laid out in this paper) apply to typical samplers used for approximate inference used by the sampling community at NeurIPS, the reviewers are excited by the potential for the novelty of this approach to inspire further work at the intersection of these areas. The authors should make sure to address the reviewers' concerns about the naming and discussion of the "non-adversariality" condition in their camera-ready version. See the updated reviews from Reviewers 1 and 2. In particular, the authors should be sure to rename this condition and discuss its restrictions more fully in their final paper.