Geometric Learning for Infinite-Dimensional Signals
Background: Deep learning architectures often operate on signals supported on manifolds, but current methods struggle with infinite-dimensional signals such as time series or probability distributions. Existing geometric learning frameworks generally assume finite-dimensional fiber representations and rely on fixed connections.
Question / Future Work: Establishing a unified theory and convolutional framework for operating directly on infinite-dimensional signals supported on manifolds, such as sections of a Hilbert bundle, is needed to ensure robustness to discretization and scalability. Extending convergence results for discrete graph-based Laplacians to the infinite-dimensional Hilbert bundle setting remains a major theoretical bottleneck for ensuring consistent performance across varying data resolutions.
Why It Matters: Addressing this gap is necessary to provide theoretical guarantees, such as consistency and transferability, for deep learning models operating on complex data structures like time series and measure-valued representations.
Evidence: The main technical reason behind this gap is the lack of an extension of the convergence result… to bundles with infinite-dimensional fibers.
Metadata & Links
- created_at
- 2026-05-10T06:05:59Z