NeurIPS 2020

Optimal Variance Control of the Score-Function Gradient Estimator for Importance-Weighted Bounds

Meta Review

This paper is accepted, however, there are a few issues that require revision. All reviewers agreed that OVIS is a novel control variate that resolves an important open problem. However, the use of the IWR bound muddled the experimental conclusions and lead to confusion. Carefully separating out the effects of OVIS and using the IWR bound is critical for improving this paper. R1 was concerned about the sensitivity to the annealing scheme and the annealing details. Furthermore, this made it challenging to compare methods in some experiments. Figure 2 in the rebuttal is a step in the right direction, and based on that effort, I believe that the authors can clarify the main paper. R1 asked about numerical instabilities and the authors explained that they clipped the importance weights. This introduces bias and potentially complicates experimental conclusions. Given that the expression in question appears in VIMCO and can be computed numerically stably without clipping, I ask that the authors rerun their experiments without clipping or provide evidence that the bias from clipping is limited. R1 asks about the computational complexity of the method. While, we agree that the method requires O(K+S) samples and expensive function evaluations, evaluating the estimator potentially requires O(KS) operations. The authors should explain how to compute this in O(K+S) operations or clarify this. While the reviewers found the paper clear, I did not find the notation clear. The authors should carefully balance defining new variables versus writing them out.