A new framework lets one blueprint build neural network layers for ten different curved-space geometries instead of designing each one from scratch.
The paper, "Building Transformation Layers for Riemannian Neural Networks" (arXiv:2609.35436), proposes a general method for building fully connected and convolutional layers on Riemannian manifolds - curved spaces where standard Euclidean math does not directly apply. Earlier work solved this problem one manifold at a time, each requiring its own specialized construction. The new framework instead works across any "computationally tractable" Riemannian space, and the authors demonstrate it on ten manifolds: three hyperbolic models, five variants of the symmetric positive definite (SPD) manifold, and two Grassmannian formulations. Code for the framework, called RieTrans, is available on GitHub.
Hyperbolic layers already power hierarchical and tree-like data models, SPD layers show up in EEG and diffusion-MRI analysis, and Grassmannian layers handle subspace and action-recognition tasks - but until now, switching between them meant rebuilding the math from the ground up. Collapsing ten bespoke designs into one general pattern matters less for any single application and more for the researchers who won't have to reinvent geometric deep learning every time they pick a new manifold.
Tractable is doing a lot of work in that phrase, though, and ten manifolds is still a finite list - the real test is whether this framework absorbs the next weird geometry someone proposes, not just the ones it already knows.