Research map / Graph learning / GNN architectures
Beyond ReLU: Bifurcation, Oversmoothing, and Topological Priors
Erkan Turan, Gaspard Abel, Maysam Behmanesh, Emery Pierson, Maks Ovsjanikov
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 |
|---|---|---|---|
| replace ReLU with sine/tanh activation and bifurcation-aware supercritical initialization (ablation) | deep GCN (64 layers) · Cora | accuracy / Dirichlet energy | Improved |
| initialize in supercritical regime (delta > 0) instead of subcritical (ablation) | 64-layer GNN with sin/tanh · Cora | accuracy | Improved |
| supercritical initialization with ReLU (breaking contractivity only) (ablation) | 64-layer GNN with ReLU · Cora | accuracy | No clear effect |
| add polynomial spectral filter to sine GCN (ablation) | Sine GCN · benchmark average rank | mean rank: 10.73 -> 3.00 | Improved |
| replace ReLU with sine activation in Poly(A) models (ablation) | Poly(A) GCN · benchmark average rank | mean rank: 6.91 -> 3.00 | Improved |
| replace sine with GELU or Swish in Poly(A) models (ablation) | Poly(A) GCN · benchmark average rank | mean rank: 3.00 -> 4.82 (GELU-pl), 7.36 (Swish-pl) | Hurt |
| use sine instead of ReLU with polynomial filter (ablation) | Poly(A) GCN · Squirrel | accuracy (%): 53.57 -> 60.81 | Improved |
| use sine instead of ReLU with polynomial filter (ablation) | Poly(A) GCN · Texas | accuracy (%): 92.79 -> 85.59 | Hurt |
| sine activation with polynomial filter vs GCN on shallow benchmarks | GCN · Photo | accuracy (%): 91.2 -> 95.69 | Improved |
| learnable polynomial filter selects topological mode | sine GNN · 10x10 grid | pattern type | Improved |
Closest papers by meaning
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · similarity 0.88)
- A Dynamical Systems-Inspired Pruning Strategy for Addressing Oversmoothing in Graph Attention Networks (ICML 2025 · similarity 0.88)
- Graph Neural Networks Do Not Always Oversmooth (NeurIPS 2024 · similarity 0.87)
- A Signed Graph Approach to Understanding and Mitigating Oversmoothing (NeurIPS 2025 · similarity 0.87)
- On Oversquashing in Graph Neural Networks Through the Lens of Dynamical Systems (AAAI 2025 · similarity 0.87)
- Let's do the time-warp-attend: Learning topological invariants of dynamical systems (ICLR 2024 · similarity 0.86)
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.
