Many scientific datasets are naturally defined on a sphere, such as cosmic microwave background maps, and have no preferred orientation. Standard convolutional networks treat the sphere as a flat grid, ignoring this symmetry, which forces researchers to rely on expensive simulations to compensate. The new architecture from Williamson, Jeffrey, and Whiteway addresses this by embedding SO(3) rotational symmetry directly into the network.
The network is built for the HEALPix pixelisation, using steerable filters, parallel transport, and spin-weighted features to construct convolutions that are provably equivariant up to a band-limit. Because the symmetry is exact in the architecture, the model does not need to learn it from data. The authors report per-cent-level or lower SO(3)-invariance error, controlled by hyperparameters, and the most accurate dense gradient regression on benchmarks, with competitive performance on CMB foreground removal and denoising.
The main payoff is in simulation-based inference, where high-fidelity simulations are costly and neural compressors need large training sets. By respecting the symmetry, the method reduces the number of simulations needed. In a cosmology lensing SBI test, it produced the tightest posteriors at a matched simulation count, suggesting that fixed simulation budgets can yield more accurate scientific results.