NeurIPS 2020

Assessing SATNet's Ability to Solve the Symbol Grounding Problem

Meta Review

This paper is a critical follow-up to the SATNet paper at ICLR last year which introduced a differentiable MAXSAT solver (and won a best paper award). The paper is somewhat critcal of some of the claims of the SATNet paper and proposes some "tricks" for overcoming these problems. Strengths identified by the reviewers include that the improvements to SATNet will be practically useful and that the analysis and critiques are insightful. Weaknesses are that some of the claimed problems with SATNet seem to be more a misunderstanding of or ambiguity in what was claimed in that paper, as opposed to fundamental issues. The authors addressed these issues in the rebuttal and promise to tone down the claims. The AC also read the paper and agrees that the paper should be accepted but authors must make the claims and framing clearer in the camera ready.