spherical cnns
**Spherical CNNs** are **neural networks that perform convolution directly on the surface of a sphere ($S^2$) rather than on flat image planes** — using either spectral methods (spherical harmonic transforms) or spatial methods (icosahedral discretization) to achieve rotation equivariance ($SO(3)$) without the distortion artifacts inherent in projecting spherical data onto flat 2D grids.
**What Are Spherical CNNs?**
- **Definition**: Spherical CNNs (Cohen et al., 2018; Esteves et al., 2018) generalize the concept of planar convolution to the sphere by defining convolution as a correlation operation over the rotation group $SO(3)$. Just as planar convolution slides a filter across a 2D grid (translation group), spherical convolution rotates a filter across all orientations on the sphere (rotation group), producing a feature map defined on $SO(3)$.
- **Spectral Approach**: The most mathematically elegant approach computes spherical convolution in the spectral domain using Spherical Harmonic Transforms (SHT) — the spherical analog of the Fourier transform. Convolution becomes pointwise multiplication in the spectral domain: $hat{f} cdot hat{g}$ where $hat{f}$ and $hat{g}$ are the spherical harmonic coefficients. This approach achieves exact $SO(3)$-equivariance but requires careful handling of bandwidth and aliasing.
- **Spatial Approach**: The alternative is to discretize the sphere using a mesh (typically an icosahedron refined to desired resolution) and define convolution through local patch operations on the mesh. This approach is more computationally tractable for high resolutions but provides only approximate equivariance depending on the mesh symmetry.
**Why Spherical CNNs Matter**
- **Omnidirectional Vision**: 360° cameras, LiDAR point clouds projected onto range spheres, and panoramic imagery all produce spherical data that is severely distorted by equirectangular projection. Spherical CNNs process this data natively on the sphere, eliminating pole distortion and the resolution waste of over-sampling near the poles.
- **Global Climate and Weather**: Earth observation data — satellite imagery, atmospheric measurements, ocean temperature fields — is fundamentally spherical. Planar CNNs applied to map projections produce systematic errors near the poles and across projection boundaries. Spherical CNNs provide rotation-equivariant analysis of global data without projection artifacts.
- **Molecular Shape Analysis**: Molecular surfaces and electron density distributions are naturally represented as functions on the sphere centered at each atom. Spherical CNNs enable rotation-equivariant analysis of molecular shape, electrostatic potential, and binding pocket geometry — critical for computational drug design.
- **Cosmology**: The Cosmic Microwave Background (CMB) is a signal measured on the celestial sphere. Spherical CNNs provide the natural architecture for analyzing CMB anisotropy patterns, searching for statistical anomalies, and testing cosmological models using full-sky data.
**Spherical CNN Approaches**
| Approach | Method | Key Trade-off |
|----------|--------|---------------|
| **Spectral (SHT)** | Convolution via spherical harmonic transform | Exact equivariance, expensive for high bandwidth |
| **Icosahedral** | Mesh-based convolution on icosahedral grid | Scalable resolution, approximate equivariance |
| **HEALPix** | Hierarchical Equal Area pixelization | Equal-area sampling, used in cosmology |
| **Equirectangular + Padding** | Standard CNN with circular padding | Simple but distorted at poles |
**Spherical CNNs** are **globe-trotting AI** — vision systems that process the world as a sphere rather than a flat map, eliminating the systematic distortions of 2D projection and enabling rotation-equivariant analysis of omnidirectional, planetary, and molecular data.