Equivariant Geodesic Networks (EGN) & egnlib
Published in Research code — Equivariant Geodesic Networks (AAAI 2026, under review), 2025
Equivariant Geodesic Networks (EGN) is a geometry-aware architecture that learns directly on the symmetric positive definite (SPD) manifold rather than projecting features to Euclidean space. It combines equivariant mapping, Riemannian pooling, a geometric bias term, and geodesic attention, with a numerically stable matrix-logarithm backward pass.
The full architecture is released as the open-source Python package egnlib (pip install egnlib). The accompanying paper is under review at AAAI 2026.