NeurIPS 2020

Learning Invariants through Soft Unification


Meta Review

Pros: - Topic is interesting: learning something akin to symbolic reasoning - Nice idea of Identifying invariants through soft-attention in expressions via task- specific examples - Informative rebuttal addressing many of the points raise - Paper well written Cons: - Missing some analysis to understand what the model is doing (show with ground truth the success in identifying variables) After reading the author’s response, R3 decided to increase his/her score to marginally above acceptance threshold, all other reviewers maintained their original scores. R3 finds the author’s response satisfactory, but would like to see analysis over learning rates. In particular, he/she considers it useful to include training curves. R2 recommends to reject the paper considering that the authors should provide a quantitative-ablative analysis of the various components in the framework to substantiate their claims. And considers it to be essential to show how well the model can correctly identify the variables (at least in the synthetic case). While R4 agrees with the points raised by R3 and R2, he/she argues that the novelty of the approach (and relevance of the problem) renders the paper particularly interesting. The AC agrees with this view. The AC considers the current results informative and given that the authors promised to report a quantitative analysis of identifying invariants, recommends accepting the paper. The AC also encourages the authors to report training curves as mentioned by R3.