NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:7112
Title:Minimum Stein Discrepancy Estimators


		
The authors introduce a general framework using kernels stein discrepancy (KSD) for estimation, which includes many existing estimators as a special case, and clearly demonstrates that such generalized discrepancies based on KSD are better at estimating non-smooth distributions and heavy-tailed distributions compared to the classical score matching (SM). This improvement is valuable to the community. There is a consensus among the reviewers that the paper should be accepted. Hence, I recommend this paper for publication at NeurIPS2019. Below are some important points that would improve the paper further. I urge the authors to take them into account for the camera-ready version. - provide more extensive experiments. - one of the reviewer complains that the paper is (unnecessarily) technically dense, which I concur. Please improve this aspect of the paper. - Please also incorporate the explanations provided in the rebuttal to the camera-ready version.