NeurIPS 2020

Learning Disentangled Representations and Group Structure of Dynamical Environments

Meta Review

4 knowledgeable reviewers reviewed the paper. Requested clarifications and points raised by the reviewers, s.a. initially missed related work discussions, were adequately addressed in the rebuttal. As a result two reviewers increased their scores, 3 voting to accept the paper, while R2 increased his score by 1 still judging the paper marginally below threshold and requiring further work. The main weakness remains the limited empirical validation, limited to toyish data, and in comparison to a limited range of baselines. All reviewers however agreed that the paper provides a novel, original and interesting contribution to the topic of disentangled representations learning. Thus, despite the mentioned limitations, and in agreement with the majority of reviewers, the AC recommends to accept the paper. We expect the clarifications form the authors' response to be integrated in the final paper. Also authors should include a more thorough discussion of scalability considerations of the method, and an explicit discussion of its limitations (as requested by R2).