NeurIPS 2020

Learning Graph Structure With A Finite-State Automaton Layer

Meta Review

In graph nets, edges can represent two kinds of relations: ones that follow immediately from the structure of the graph, and ones that are abstract/implicit. The paper proposes to learn the latter. More precisely, it considers relations defined as paths in the base graph accepted by a finite-state automaton, poses the problem of learning these relations as a POMDP problem, and solves a relaxed version of this problem using gradient descent. Overall, the paper was well-received. Pros: + Fresh idea + Clean formulation + Experiments show clear gains in the domains considered + The paper is well-written Cons: - Some missing related work - Somewhat narrow application domain The reviewers appreciated the clarifications provided in the author response, in particular the RL experiment for the "Go for a Walk" domain. Please integrate your responses in the rebuttal with the main paper. And naturally, consult the reviews for more detailed feedback.