Review for NeurIPS paper: PLANS: Neuro-Symbolic Program Learning from Videos

Relation to Prior Work: The relation to Ellis 2018 (which the authors discuss) should be reframed. They also learn to infer specifications from noisy perceptual input, which are then fed to a downstream symbolic solver, and also addresses the challenge of uncertainty over specifications, albeit in a Bayesian way rather than via the heuristics proposed here. Could you similarly situate your system in a probabilistic framework, and resolve the ambiguity over specs in a less heuristic manner? Would that fare better or worse on your data sets? I feel this is the main substantive difference, rather than the details which are presently emphasized in the text.