Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Jacobsen, Albert Kjøller, Jakobsen, Leo Uhre, Gegenfurtner, Johanna Marie, Arvanitidis, Georgios
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2605.15459
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866916015365947392
author Jacobsen, Albert Kjøller
Jakobsen, Leo Uhre
Gegenfurtner, Johanna Marie
Arvanitidis, Georgios
author_facet Jacobsen, Albert Kjøller
Jakobsen, Leo Uhre
Gegenfurtner, Johanna Marie
Arvanitidis, Georgios
contents The minima of modern neural network loss functions are typically not isolated, rather they form connected components of reparameterization invariant solutions on the training data. Analytically characterizing these solutions is a hard problem, but sampling approaches are feasible. By construction, existing methods either spread over low-loss regions, and thus do not sample reparameterization invariant solutions exactly, or are inherently local, which limits exploration of other minima valleys. We propose sampling such reparameterization invariant models using a dynamical system based on kinetic energy, subject to a gravitational pull and a friction term that dissipates energy from the system. Our proposed sampler, DiMS, is guaranteed to sample exactly from the minimum level sets and depends on physically motivated hyperparameters which allows control over the exploration capabilities of the sampler. We consider uncertainty quantification in Bayesian inference as the motivating problem and observe improved performance compared to previously proposed approaches.
format Preprint
id arxiv_https___arxiv_org_abs_2605_15459
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Don't Stop Me Yet: Sampling Loss Minima via Dissipative Riemannian Mechanics
Jacobsen, Albert Kjøller
Jakobsen, Leo Uhre
Gegenfurtner, Johanna Marie
Arvanitidis, Georgios
Machine Learning
The minima of modern neural network loss functions are typically not isolated, rather they form connected components of reparameterization invariant solutions on the training data. Analytically characterizing these solutions is a hard problem, but sampling approaches are feasible. By construction, existing methods either spread over low-loss regions, and thus do not sample reparameterization invariant solutions exactly, or are inherently local, which limits exploration of other minima valleys. We propose sampling such reparameterization invariant models using a dynamical system based on kinetic energy, subject to a gravitational pull and a friction term that dissipates energy from the system. Our proposed sampler, DiMS, is guaranteed to sample exactly from the minimum level sets and depends on physically motivated hyperparameters which allows control over the exploration capabilities of the sampler. We consider uncertainty quantification in Bayesian inference as the motivating problem and observe improved performance compared to previously proposed approaches.
title Don't Stop Me Yet: Sampling Loss Minima via Dissipative Riemannian Mechanics
topic Machine Learning
url https://arxiv.org/abs/2605.15459