Learning Directional Randers Geometry for Latent-Space Exploration
| Source: arXiv AI
Tags: Finsler geometry, Randers metric, metric learning, latent space, Alzheimer's disease, GE Healthcare
A new Finsler-Randers metric learning framework gives neural latent spaces direction-dependent distances, so interpolations can respect one-way processes like disease progression.
Details
Latent spaces from deep networks lack explicit geometry, and standard interpolation assumes moving from A to B costs the same as B to A. That breaks for directional processes such as tumor growth or cognitive decline in Alzheimer's disease. Riemannian metrics add curvature but stay symmetric. The authors argue asymmetry, not curvature alone, is the missing piece, and propose learning a Finsler-Randers metric whose distances depend on direction. This lets the model compute trajectories that respect monotonic sequential constraints. The work, from HeKA/GE Healthcare and CMAP, demonstrates interpolation and extrapolation on controlled cross-sectional datasets rather than large real-world benchmarks.