Research map / Architectures and efficiency / CNN design
Platonic Transformers: A Solid Choice For Equivariance
Mohammad Mohaiminul Islam, Rishabh Anand, David R Wessels, Friso de Kruiff, Thijs P. Kuipers, Rex Ying, Clara I. Sánchez, Sharvaree Vadgama, Georg Bökman, Erik J Bekkers
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 |
|---|---|---|---|
| add 2D rotational group weight-sharing (D3) to the ViT (ablation) | ViT with RoPE · CIFAR-10 | accuracy (attention): 91.65 -> 92.78 | Improved |
| use a larger symmetry group (C6/D3) with fewer parameters (ablation) | Platonic Transformer · CIFAR-10 | accuracy: 14.2M params vs 85.1M for baseline | Improved |
| keep softmax attention instead of linear dynamic convolution (ablation) | Platonic Transformer · CIFAR-10 | accuracy: +5.31 percentage points averaged over groups | Improved |
| add Tetrahedron group weight-sharing (ablation) | ViT with RoPE · ScanObjectNN | overall accuracy (attention): 83.68 -> 86.11 | Improved |
| keep softmax attention instead of linear dynamic convolution (ablation) | Platonic Transformer · ScanObjectNN | overall accuracy: +6.26 percentage points averaged over groups | Improved |
| use equivariant Tetrahedron group instead of trivial group (ablation) | DiP diffusion transformer · QM9 generation | molecule stability: 91.54 -> 95.26 | Improved |
| use equivariant Tetrahedron group instead of trivial group (ablation) | DiP diffusion transformer · QM9 generation | validity (%): 97.09 -> 98.43 | Improved |
| use Tetrahedron group instead of trivial group (ablation) | Platonic Transformer · ProteinMD backbone level | force MSE: 2.11 -> 1.80 | Improved |
| use Tetrahedron group instead of trivial group (ablation) | Platonic Transformer · ProteinMD atom level | force MSE: 2.60 -> 2.36 | Improved |
| increase symmetry group from Tetrahedron to Octahedron at halved parameters (30M) (ablation) | Platonic Transformer · OMol25 4M (20 epochs) | force MAE: 11.15 -> 12.28 | Hurt |
| scale Octahedron model back to 60M parameters (ablation) | Platonic Transformer Octahedron · OMol25 4M (20 epochs) | force MAE: 12.28 -> 10.51 | Improved |
| train Tetrahedron model longer (20 to 80 epochs) | Platonic Transformer Tetrahedron · OMol25 4M | force MAE: 11.15 -> 9.33 | Improved |
Closest papers by meaning
- A Program to Build E(N)-Equivariant Steerable CNNs (ICLR 2022 · similarity 0.89)
- Generalizing Convolutional Neural Networks for Equivariance to Lie Groups on Arbitrary Continuous Data (ICML 2020 · similarity 0.88)
- Probing Equivariance and Symmetry Breaking in Convolutional Networks (NeurIPS 2025 · similarity 0.86)
- On Making $SE(2)$-Invariant Networks Optimal (NeurIPS 2026 · similarity 0.86)
- Hide & Seek: Transformer Symmetries Obscure Sharpness & Riemannian Geometry Finds It (ICML 2025 · similarity 0.86)
- Neural Isometries: Taming Transformations for Equivariant ML (NeurIPS 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.
