Reproduce “Inference in Probabilistic Graphical Models by Graph Neural Networks” on OpenResearch

About this paper

KiJung Yoon, Renjie Liao, Yuwen Xiong, Lisa Zhang, Ethan Fetaya, Raquel Urtasun, Richard Zemel, Xaq Pitkow

The authors train graph neural networks to estimate variable probabilities and the most likely complete assignment in binary graphical models. Their learned updates pass information along graph connections, replacing fixed belief-propagation rules. On held-out larger and structurally different graphs, the networks often outperform belief propagation, including by more than an order of magnitude in marginal accuracy on many-loop graphs, with exceptions.

Key insights

• The iterative message-passing structure of GNNs is naturally well-suited to learn and implement probabilistic inference dynamics.
• Learning the message-passing rules with GNNs allows for adaptation to the specific characteristics of different graphical models, enabling better performance than fixed, hand-designed algorithms on loopy graphs.
• GNNs trained on a diverse set of small graphical models can generalize their learned inference capabilities to unseen graphs of different sizes and entirely novel topological structures.

Summary by alphaXiv.