NeurIPS 2020

Factor Graph Neural Networks

Meta Review

This paper generalizes graph neural networks to factor graphs, and shows that the max-product algorithm is a special case (though there can be exponential number of rank-1 tensors). Experiments synthetic data, LDPC decoding, and human motion prediction are presented. The major concerns were addressed in the rebuttal, which provides additional clarifications and evidence. The authors should include them in the final version of the paper.