Abstract
We introduce two novel constructions in geometric deep learning designed to address fundamental challenges in processing manifold-valued data. First, we propose a method for transporting orientation-dependent convolutional filters over a manifold using horizontal frame bundle flows, naturally incorporating the rotational effects of holonomy. Second, we introduce the Weighted Diffusion Mean (wDM), which allows for efficient evaluation of manifold convolution layers by sampling manifold-valued random variables via non-linear bridge sampling. These methods, rooted in stochastic analysis and geometric statistics, provide a theoretically grounded framework for building robust, equivariant, and computationally efficient neural networks on curved surfaces.
Introduction
Geometric Deep Learning (GDL) extends the success of neural networks to non-Euclidean domains such as manifolds (e.g., spheres, 3D shapes, and medical imaging data). However, two primary "geometry-induced" challenges persist: the orientation problem caused by holonomy and the high computational cost of generating manifold-valued outputs. This paper bridges stochastic analysis on manifolds with GDL to provide solutions using horizontal frame bundle flows and stochastic bridge sampling.
Addressing the Orientation Problem: Horizontal Frame Bundle Flows
In Euclidean space, orientation is globally consistent; "up" remains "up" regardless of the path taken. On a curved manifold, moving a filter from point A to point B along different paths results in different rotations—a phenomenon known as holonomy. Standard GDL methods often rely on parallel transport along the shortest path (geodesic), but this approach becomes discontinuous at the "cut locus" where multiple shortest paths exist.
To resolve this, we "lift" the problem to the orthonormal frame bundle ($OM$), the space of all possible orientations at every point. We utilize the Eells-Elworthy-Malliavin (EEM) construction of Brownian motion to define a convolution that naturally distributes orientations across the manifold. This "stochastic transport" accounts for curvature globally and avoids the topological discontinuities inherent in deterministic local methods.
By using Brownian motion on the frame bundle, we create a notion of "global" parallel transport. This construction ensures that the resulting convolution is gauge equivariant, meaning the network's output remains consistent regardless of the local coordinate system chosen.
Efficient Manifold-Valued Outputs: The Weighted Diffusion Mean
When a network layer must output a point on a manifold (such as a 3D rotation or a point on a brain surface), simple weighted averaging is insufficient. Traditionally, one computes a Fréchet mean, which requires expensive, iterative optimization at every layer—a significant bottleneck for deep learning architectures.
We propose the Weighted Diffusion Mean (wDM) as a computationally efficient alternative. Instead of solving a complex optimization problem to find a single deterministic mean, we use stochastic bridge sampling. This technique simulates random paths conditioned to hit a specific target, allowing the network to sample from a distribution centered at the mean.
This approach transforms the output problem from an optimization task into a sampling task, which can be evaluated efficiently using importance sampling (Algorithm 2 in the paper). This stochasticity acts as a feature, creating operators that are smoother and more globally aware than their deterministic counterparts.
Theoretical Insights and Properties
Our analysis yields several key theoretical results:
- Curvature and Non-Commutativity: We prove that the lack of associativity and commutativity in manifold convolutions is directly linked to the Riemannian curvature tensor. Curvature acts as a source of non-commutativity that must be accounted for by the network architecture.
- Smoothness: Unlike standard parallel transport, which "jumps" at the cut locus, the proposed stochastic flow results in a smooth integrand for analytic manifolds. This smoothness is critical for stable gradient-based learning.
- Equivariance: We demonstrate that horizontal convolutions are equivariant to the choice of local coordinates, a necessary property for consistent geometric processing.
Conclusion and Future Work
This work provides a rigorous mathematical foundation for integrating stochastic analysis into geometric deep learning. By embracing stochasticity rather than avoiding it, we circumvent the topological and computational hurdles of holonomy and Fréchet means. While this paper is primarily theoretical, it provides concrete algorithmic sketches for evaluating these complex stochastic operators. Future work will focus on empirical validation in frameworks like PyTorch and applying these methods to discrete geometries such as graphs and meshes using discrete connections.
Summary updated