NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:5764
Title:Fast-rate PAC-Bayes Generalization Bounds via Shifted Rademacher Processes


		
The paper further develops the direction of proving PAC-Bayesian bounds via techniques to control the Rademacher complexity of the class of distributions with bounded Kullback-Leibler divergence. Connecting these two important tools is of fundamental interest, and the paper has the added benefit of reminding the PAC-Bayesian community of the earlier work in this direction by Kakade, Sridharan and Tewari (2009). For the final version, I would encourage the authors to try to improve the comparisons to related work by carefully considering the remarks of the reviewers. It would also be interesting if they could reflect on whether their results might also be provable using Talagrand's inequality + peeling, as remarked by reviewer #1.