NeurIPS 2020

Deep Direct Likelihood Knockoffs


Meta Review

Three referees support acceptance for the contributions, notably a new approach to generating knockoffs based on neural networks that outperforms existing methods for generating knockoffs. R4 dissents, citing the clarity of the paper as is its more significant weakness. I disagree. In fact, this is one of the most clearly written neurips submission in my stack. Nonetheless, I encourage the authors to clarify all the ambiguous mathematical notation R4 flags. R4 also takes issue with the comparison to KnockoffGAN, which experienced mode collapse during training, and hence performed poorly. The author response, however, reveals that the authors used a reference implementation of KnockoffGAN and grid searched for hyperparameters. I commend the authors on their thoroughness here, and recommend acceptance. This submission "checks all the boxes": a novel method that is nonetheless straightforward enough to work in practice, a clear write-up, and compelling experimental results on both synthetic and real data. I encourage the authors to consider revising their manuscript to address R3's point about global nulls.