NeurIPS 2020

PEP: Parameter Ensembling by Perturbation


Meta Review

Three knowledgeable referees support acceptance for the contributions, given the simplicity, effectiveness and theoretical soundness of the proposed approach. However, please consider revising the paper to address R1 and R2's remark on - highlighting the connection with prior work on Baysian inference such as Ritter et al. 2018. - comparison with stochastic based ensembling approach such as Stochastic Weight Averaging (SWA). - Evaluation results on out-of-distribution detection tasks.