Gaussian-Smoothed Sliced Probability Divergences

Gaussian smoothed sliced Wasserstein distance has been recently introduced for comparing probability distributions, while preserving privacy on the data. It has been shown that it provides performances similar to its non-smoothed (non-private) counterpart. However, the computational and statistical properties of such a metric have not yet been well-established. This work investigates the theoretical properties of this distance as well as those of generalized versions denoted as Gaussian-smoothed sliced divergences G σ SD p. We first show that smoothing and slicing preserve the metric property and the weak topology. To study the sample complexity of such divergences, we then introduce μn the double empirical distribution for the smoothed-projected µ. The distribution μn is a result of a double sampling process: one from sampling according to the origin distribution µ and the second according to the convolution of the projection of µ on the unit sphere and the Gaussian smoothing. We particularly focus on the Gaussian smoothed sliced Wasserstein distance and prove that it suffers from an unavoidable bias of approximation of order log(n/2). We also derive other properties, including continuity, of different divergences with respect to the smoothing parameter. We support our theoretical findings with empirical studies in the context of privacy-preserving domain adaptation.

Domaines

Apprentissage [cs.LG] Statistiques [math.ST]

Fichier principal

gs_swd_arxiv.pdf (1.63 Mo)

cifar-div-vs-samples.png (191.51 Ko)

da-A-W.png (82.17 Ko)

da-U-M.png (89.49 Ko)

da-W-D.png (76.54 Ko)

dareg-A-W.png (64.03 Ko)

dareg-U-M.png (84.22 Ko)

dareg-W-D.png (69.29 Ko)

gs_swd_arxiv.synctex.gz (313.73 Ko)

mnist-div-vs-samples.png (191.39 Ko)

toy-approx-proj.png (217.21 Ko)

toy-div-displacement.png (238.35 Ko)

toy-div-vs-noise-0.png (133.18 Ko)

toy-div-vs-noise-2.png (129.85 Ko)

toy-div-vs-noise.png (133.18 Ko)

toy-div-vs-samples-noise.png (210.64 Ko)

toy-div-vs-samples.png (201.36 Ko)

toy-div-vs-samples_dim.png (170.47 Ko)

Origine	Fichiers produits par l'(les) auteur(s)

Origine	Fichiers produits par l'(les) auteur(s)

Mokhtar Alaya : Connectez-vous pour contacter le contributeur

https://hal.science/hal-04282819

Soumis le : lundi 13 novembre 2023-15:54:46

Dernière modification le : vendredi 5 avril 2024-03:24:56

Dates et versions

hal-04282819 , version 1 (13-11-2023)

hal-04282819 , version 2 (03-04-2024)

hal-04282819 , version 3 (22-04-2024)

Identifiants

HAL Id : hal-04282819 , version 1

Citer

Mokhtar Z. Alaya, Alain Rakotomamonjy, Maxime Berar, Gilles Gasso. Gaussian-Smoothed Sliced Probability Divergences. 2023. ⟨hal-04282819v1⟩

Exporter

BibTeX XML-TEI Dublin Core DC Terms EndNote DataCite

240 Consultations

140 Téléchargements