Crate stableprop 

Stable distribution propagation through neural network layers.

Implements moment-matching propagation of Gaussian distributions through affine (linear) and ReLU layers. The linear case is exact; the ReLU case uses the Frey & Hinton (1999) moment-matching approximation that computes post-ReLU mean and variance from the Gaussian CDF and PDF evaluated at mu / sigma.

The ReLU step is the Frey & Hinton (1999) Gaussian moment-matching approximation, with off-diagonal covariance dropped (diagonal assumption). Keeping the full covariance and the heavy-tailed (Cauchy) case is the generalization of Petersen et al., “Uncertainty Quantification via Stable Distribution Propagation” (ICLR 2024), which this crate does not implement.

The burn_sdp module (feature burn) provides the same propagation on Burn tensors: batched, differentiable, and composable with Burn models.

Modules

burn_sdp

Distribution propagation on Burn tensors (diagonal-Gaussian / assumed-density filtering).

Structs

Moments

First two moments of a multivariate Gaussian (mean + full covariance).

Enums

Layer

A single neural-network layer.

Functions

propagate_linear

Propagate Gaussian moments through an affine (linear) layer.

propagate_relu

Propagate Gaussian moments through an element-wise ReLU using Frey & Hinton (1999) moment matching.

propagate_sequential

Propagate moments through a sequence of layers.