NeurIPS 2020

CASTLE: Regularization via Auxiliary Causal Graph Discovery

Meta Review

This paper proposes a new regularization method based on learning causal graphs. Three reviewers suggest accept, one indicates reject. All reviewers found that the method was extensively compared against proper baselines. The concerns about scalability were addressed by the rebuttal. R3 recommended reject based on concerns around correctness, which I found were adequately addressed in the rebuttal. Therefore, I recommend accept.