Bounded Random Features for Grassmannian Kernels with Rank-One Projections
Rémi Delogne (PhD Student, UClouvain)
| When |
Nov 25, 2025
from 01:00 to 02:00 |
|---|---|
| Where | M7 101 |
| Attendees |
Rémi Delogne |
| Add event to calendar |
vCal
iCal
|
Rémi Delogne
Title: Bounded Random Features for Grassmannian Kernels with Rank-One Projections
Abstract:
We develop scalable random feature approximations for Grassmannian kernels using asymmetric rank-one projections (AROP). While AROP provides an unbiased sketch of projection matrices, its heavy-tailed behaviour prevents reliable concentration. We address this by applying bounded transforms—binary (sign) and complex exponential—which yield positive-definite Grassmannian kernels with sub-Gaussian concentration and, in the exponential case, an explicit closed form in principal angles.
We further introduce structured AROP sketches based on fast transforms, reducing storage and computation by orders of magnitude while preserving statistical accuracy. Experiments on ETH-80 show that the proposed random features closely match exact Grassmannian kernels, and that structured variants recover accuracy efficiently. This provides a practical framework for fast, geometry-aware learning on the Grassmann manifold.
Website: https://scholar.google.com/citations?user=2M2g8_AAAAAJ&hl=en
In Room M7 101, 1st floor, Monod campus, ENSL.