Frequency-subspace margin decomposition

methods / Theoretical / Analytical / Geometric analysis / Frequency-subspace margin decomposition

Techniqueadvanced

Decomposes a classifier's decision-boundary margin (distance to the boundary via subspace-constrained DeepFool) across an orthogonal DCT-frequency basis rather than reporting a single aggregate margin, revealing that a network's invariance/sensitivity to perturbation is highly direction-dependent rather than isotropic.

Used in (1 observation)

Decision-boundary margin is small along low-frequency discriminative directionsmeasured

structure: Decision boundary (as a codimension-1 hypersurface) · models: LeNet (image classifier, MNIST/Fashion-MNIST-scale), DenseNet-121, ResNet-50 (supervised, ImageNet), Overparameterized MLP on synthetic dataset T_1 · paper: Hold Me Tight! Influence of Discriminative Features on Deep Network Boundaries