Graph Neural Network

A neural network built for graph-structured data, where each node repeatedly exchanges information with its neighbors along edges.

A graph neural network processes data made of nodes and edges; its core operation is message passing: at each layer, every node aggregates information from its neighbors to update its own features, and stacking multiple layers lets information reach further-away nodes, with the result independent of how the nodes happen to be numbered. Scarselli and colleagues proposed the 'graph neural network model' in 2009, followed by variants such as graph convolutional networks (GCN) and graph attention networks (GAT). Embodied AI has three common uses for it: representing a robot's own body as a graph, with joints and limbs as nodes — as in NerveNet, which lets a policy transfer across agents with different body shapes; representing particles or objects as a graph to learn physics, as in DeepMind's 2020 Graph Network Simulator (GNS); and representing object relationships in a 3D scene graph for reasoning and planning.

ExampleGNS treats every particle of a material like fluid or sand as a node, connects nearby particles with edges, and predicts each particle's next position through several rounds of message passing; at test time it can handle an order of magnitude more particles than it was trained with.

Also called

GNN, Graph Network, Message-Passing Neural Network

Related

3D Scene Graph · Neural Simulator · Point Cloud Encoder · Transformer · Morphology-Control Co-design · Inductive Bias

Sources

Graph neural network - Wikipedia
Learning to Simulate Complex Physics with Graph Networks (Sanchez-Gonzalez et al., ICML 2020)
NerveNet: Learning Structured Policy with Graph Neural Networks (project page)