Skip to main content
Search

# Generalized Gauss-Newton Approximation

1 related articles

Subscribe to updates
  1. arXiv

    Generalized Laplace Active Subspaces: Scalable Neural-Network Uncertainty Quantification with a Handful of Curvature Directions

    The work proposes a low-rank Laplace approximation for neural-network uncertainty quantification within a generalized Bayesian framework, restricting the posterior to an active subspace spanned by dominant generalized Gauss-Newton curvature directions of the empirical mean loss, computing eigenpairs matrix-free via Lanczos and calibrating the prior variance by empirical Bayes; across regression and molecular-generation classification tasks, standard Bayesian scaling contracts posterior variance with data size and forces retention of many weak-curvature directions, whereas mean-loss scaling yields calibrated, coherent predictive intervals with only a few directions.

You’re all caught up