NeurIPS 2020

Learning Physical Constraints with Neural Projections

Meta Review

The paper proposes an approach for modeling dynamical systems. Instead of directly modeling the transition function it learns a linear forward mapping together with a constraint satisfaction function that can be used as a learned neural projection operator. The reviewers agree that this an interesting, non-obvious, and novel idea that could stimulate considerably future work. The reviewers also raised several concerns in particular with respect to the quality of the evaluation which in the form presented in the submission is not entirely satisfying (scope of the experiments; quality of the baselines). Overall, however, the reviewers were satisfied with the author response and recommend acceptance. We strongly encourage the authors to take into account the reviewers’ comments and improve the empirical evaluation and to be more upfront about the limitations of the proposed approach in the final version.