Principal curvature estimation (MAPC)

methods / Theoretical / Analytical / Geometric analysis / Principal curvature estimation (MAPC)

Techniqueadvanced

Estimates the mean absolute principal curvature (MAPC) of the local manifold traced out by a layer's activations, quantifying how much the representation manifold bends away from flat (affine) at each layer, distinct from intrinsic dimension (which measures how many coordinates are needed, not how curved the manifold is).

Used in (1 observation)

Latent-manifold principal curvature (MAPC) tracks CNN generalizationmeasured

structure: Curvature profile of the representation manifold · models: ResNet (image classifier, various depths), VGG (image classifier, various depths) · paper: Data Representations' Study of Latent Image Manifolds