arXiv Statistics ML By Marco Miani

Calibrating your confidence requires thinking of all the ways you could be wrong

Read the original on arXiv Statistics ML →

The paper introduces a novel Bayesian posterior—a degenerate Gaussian with covariance aligned to the null space of the Fisher information matrix—to address deep learning’s confidence miscalibration. Two efficient algorithms enable scalable posterior sampling and direct predictive confidence computation, achieving state‑of‑the‑art calibration and out‑of‑distribution detection on large models. The work also provides theoretical insights via information geometry, corrects Laplace approximation misalignments, and connects to PAC‑Bayes generalization bounds.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Statistics ML.

arXiv AI
Oct 9

Scalable AI Uncertainty Quantification via Generalized Laplace Active Subspaces

The paper introduces a scalable method for uncertainty quantification in neural networks by applying a low‑rank generalized Laplace approximation within a data‑informed active subspace. It constructs a local Gaussian posterior around pretrained weights, calibrates the prior variance via empirical Bayes, and compares two posterior scalings: standard Bayesian and mean‑loss scaling. The study finds that mean‑loss scaling yields a more stable, lower‑dimensional subspace and produces calibrated, coherent predictive confidence intervals, whereas standard scaling can cause posterior variance contraction and degrade predictive coherence.

By Wouter N. Edeling, Peter V. Coveney
arXiv Machine Learning
Jun 24

The Degeneracy Distillery

arXiv:2606. 23838v1 Announce Type: new Abstract: When two or more parameters or labels produce similar data, they are degenerate, or hard to distinguish.

By T. Lucas Makinen, Deaglan J. Bartlett, Niall Jeffrey, Benjamin D. Wandelt