Abstract
This paper presents an approach to learning (deep) $n$D features equivariantunder orthogonal transformations, utilizing hyperspheres and regular$n$-simplexes. Our main contributions are theoretical and tackle majorchallenges in geometric deep learning such as equivariance and invariance undergeometric transformations. Namely, we enrich the recently developed theory ofsteerable 3D spherical neurons -- SO(3)-equivariant filter banks based onneurons with spherical decision surfaces -- by extending said neurons to $n$D,which we call deep equivariant hyperspheres, and enabling their multi-layerconstruction. Using synthetic and real-world data in $n$D, we experimentallyverify our theoretical contributions and find that our approach is superior tothe competing methods for benchmark datasets in all but one case, additionallydemonstrating a better speed/performance trade-off in all but one other case.