H-CNNs: Convolutional Neural Networks for Riemannian Homogeneous Spaces
share
Convolutional neural networks are ubiquitous in Machine Learning applications for solving a variety of problems. They however can not be used as is when data naturally reside on commonly encountered manifolds such as the sphere, the special orthogonal group, the Grassmanian, the manifold of symmetric positive definite matrices and others. Most recently, generalization of CNNs to data residing on a sphere has been reported by some research groups, which go by several names but we will refer to them as spherical CNNs (SCNNs). The key property of SCNNs distinct from the standard CNNs is that they exhibit the rotational equivariance property. In this paper, we theoretically generalize the SCNNs to Riemannian homogeneous manifolds, that include many commonly encountered manifolds including the aforementioned example manifolds. Proof of concept experiments involving synthetic data generated on the manifold of (3 × 3) symmetric positive definite matrices and the product manifold of R^+ ×S^2 respectively, are presented. These manifolds are commonly encountered in diffusion magnetic resonance imaging, a non-invasive medical imaging modality.
READ FULL TEXT