This paper investigates the stability of parallel transport convolution, a key component in geometric deep learning, when subjected to perturbations in the underlying Riemannian metric. Using differential geometry and analysis, the research establishes that for continuous input signals and kernels, the convolution output remains continuously dependent on the Riemannian metric in the L2 space under specific conditions. The study addresses challenges posed by the cut locus by employing uniform estimates away from it and integral bounds on small neighborhoods, ultimately connecting perturbation theory for Riemannian geometric constructions with spatial convolution operators. AI
IMPACT Establishes theoretical stability guarantees for geometric convolutions, potentially informing future robust deep learning architectures on non-Euclidean data.
RANK_REASON This is a research paper published on arXiv detailing theoretical work in geometric deep learning. [lever_c_demoted from research: ic=1 ai=1.0]
- Analysis
- arXiv
- C^2-topology
- compact Lie groups
- cs.LG
- Cut locus
- differentiable pinching problems
- differential geometry
- geodesics in general relativity
- Geometric Deep Learning: Going beyond Euclidean data
- L2
- manifold
- parallel transport
- Parallel Transport Convolution: Deformable Convolutional Networks on Manifold-Structured Data
- polynomial ring in two variables
- Riemannian manifold
- Riemannian metric
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →