Research map / Graph learning / GNN architectures
How Neural Networks Extrapolate: From Feedforward to Graph Neural Networks
Keyulu Xu, Mozhi Zhang, Jingling Li, Simon Shaolei Du, Ken-Ichi Kawarabayashi, Stefanie Jegelka
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use max readout instead of sum readout (ablation) | GNN for max degree · Max degree extrapolation | MAPE | Improved |
| Use min aggregation and min readout instead of sum-based operations (ablation) | GNN for shortest path · Shortest path extrapolation | MAPE | Improved |
| Encode the force calculation in edge features (ablation) | GNN for the n-body problem · n-body problem, unseen masses or distances | MAPE | Improved |
| Train on graphs with varied rather than restricted degree structures (ablation) | GNN with max readout · Max degree extrapolation | MAPE | Improved |
| Train on graphs with extremely sparse or dense structures (ablation) | GNN with min operations · Shortest path extrapolation | MAPE | Hurt |
| Use a training distribution covering all directions rather than restricted directions (ablation) | ReLU MLP learning a linear target · Linear-function extrapolation | Extrapolation error | Improved |
| Use ReLU MLPs to extrapolate nonlinear targets that are not nearly linear along directions from the origin | ReLU MLP · Nonlinear function extrapolation | Extrapolation error | Hurt |
Closest papers by meaning
- Deep Neural Networks Tend To Extrapolate Predictably (ICLR 2024 · similarity 0.87)
- How Graph Neural Networks Learn: Lessons from Training Dynamics (ICML 2024 · similarity 0.86)
- Just One Layer Norm Guarantees Stable Extrapolation (NeurIPS 2025 · similarity 0.86)
- Graph Metanetworks for Processing Diverse Neural Architectures (ICLR 2024 · similarity 0.85)
- NeuralSolver: Learning Algorithms For Consistent and Efficient Extrapolation Across General Tasks (NeurIPS 2024 · similarity 0.85)
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
