GLUE (Geometry Linked to Untangling Efficiency)

methods / Theoretical / Analytical / Geometric analysis / GLUE (Geometry Linked to Untangling Efficiency)

GLUE (Geometry Linked to Untangling Efficiency)

Techniqueintermediate

A representational-geometry framework (Chou et al. 2025) building on manifold-capacity theory that computes three task-relevant per-class-manifold measures from penultimate-layer activation statistics — effective dimension (D_eff), effective radius (R_eff), and effective utility (Psi_eff) — related by N_crit = P*D_eff / (Psi_eff*(1 + R_eff^-2)); used as a diagnostic of untangling/separability without operationalizing a single scalar manifold capacity.

Used in (1 observation)

Penultimate-layer manifold geometry predicts OOD generalization failuremeasured

structure: Polytope (Simplex) · models: ResNet (image classifier, various depths), VGG (image classifier, various depths) · paper: Diagnosing Generalization Failures from Representational Geometry Markers