Local complexity density estimation (bias-noise perturbation)

methods / Theoretical / Analytical / Local complexity density estimation (bias-noise perturbation)

Local complexity density estimation (bias-noise perturbation)

Techniqueadvanced

Estimates the local density of a piecewise-linear network's region-boundary ('nonlinear locus') near a data distribution, without enumerating the full polyhedral complex, by perturbing each unit's bias with small noise and measuring how often the perturbation flips the unit's activation pattern near a given input.

Used in (1 observation)

Local-complexity density of the ReLU linear-region boundary drops during the terminal phase of training on real MNIST images, coinciding with increased adversarial robustnessmeasured

structure: Linear region arrangement (polyhedral tessellation of input space) · models: 4-layer MLP, 200 units/layer, on 1000 MNIST images (Patel & Montufar) · paper: On the Local Complexity of Linear Regions in Deep ReLU Networks