Research map / Architectures and efficiency / CNN design
Scaling Spherical CNNs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Phase collapse nonlinearity (vs magnitude/gated activation) (ablation) | small QM9/H spherical CNN · QM9 H | MAE (meV): 15.25 vs 16.70 (gated) and 17.01 (magnitude) | Helped |
| Spectral pooling (vs spatial pooling) (ablation) | small QM9/H spherical CNN · QM9 H | MAE (meV): 15.25 vs 16.13 | Helped |
| Spherical molecule representation using power-law pairwise features (vs prior Coulomb sphere representation) (ablation) | small QM9/H spherical CNN · QM9 H | MAE (meV): 15.25 vs 20.90 | Helped |
| Phase collapse nonlinearity added in efficiency analysis (ablation) | spherical CNN on QM9 H · QM9 H | RMSE / steps per second: -8.0% RMSE but -4.6% steps/s | Helped |
| Spectral batch norm (vs spatial) (ablation) | spherical CNN on QM9 H · QM9 H | RMSE and steps/s: +7.8% steps/s, -1.4% RMSE | Helped |
| Efficient spectral residual block (vs spatial pooling with batch norm) (ablation) | spherical CNN on QM9 H · QM9 H | RMSE and steps/s: +19.3% steps/s, -2.4% RMSE | Helped |
| Use DFT matrix multiplication instead of FFT (ablation) | spherical CNN on TPU · QM9 H efficiency | steps/s: +21.4% steps/s, 0.0 RMSE change | Helped |
| Removing Wigner Delta symmetries (compute all entries) (ablation) | spherical CNN on TPU · QM9 H efficiency | steps/s: +16.3% steps/s, 0.0 RMSE change | Helped |
| JAX rewrite tuned for TPUs vs original TensorFlow implementation (ablation) | spin-weighted spherical CNN · QM9 H efficiency | steps/s: +33.7% steps/s | Helped |
| Scaling input resolution down (256 to 64) with fixed layers (ablation) | weather forecasting spherical CNN · ERA5 6h-ahead forecasting | RMSE Z500 / T850: Z500 34.93 to 69.60; T850 0.62 to 1.08 | Hurt |
| Halving channels per layer (50%) at 256 resolution (ablation) | weather forecasting spherical CNN · ERA5 6h-ahead forecasting | RMSE Z500 / T850: Z500 34.93 to 41.34; T850 0.62 to 0.71 | Hurt |
| Separable convolutions in every other layer (fewer convolutions) (ablation) | weather forecasting spherical CNN · ERA5 6h-ahead forecasting | RMSE Z500 / T850: Z500 34.93 to 36.24; T850 0.62 to 0.65 | Hurt |
Closest papers by meaning
- Scattering Networks on the Sphere for Scalable and Rotationally Equivariant Spherical CNNs (ICLR 2022 · similarity 0.90)
- Efficient Generalized Spherical CNNs (ICLR 2021 · similarity 0.90)
- Spherical CNNs (ICLR 2018 · similarity 0.90)
- DeepSphere: a graph-based spherical CNN (ICLR 2020 · similarity 0.89)
- Attention on the Sphere (NeurIPS 2025 · similarity 0.85)
- Hybrid Neural Representations for Spherical Data (ICML 2024 · similarity 0.83)
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.