Adaptive online estimation for mixtures of ECD: a geometric approach - Optimization and learning for Data Science Accéder directement au contenu
Pré-Publication, Document De Travail Année : 2023

Adaptive online estimation for mixtures of ECD: a geometric approach

Résumé

Mixtures of elliptically-contoured distributions are highly versatile at modeling real-world probability distributions. They have therefore played a valuable role in computer vision and image processing, radar and biomedical signal processing. Existing methods for the estimation of these mixtures may become impractical for relatively-large datasets, either due to lack of computational resources or to poor performance (slow convergence or inaccuracy). To overcome these issues, the present paper introduces a new estimation method, called the CIG method (component-wise information gradient). On the one hand, this is an online method, so it requires moderate computational resources. On the other hand, it uses an adaptive step-size selection rule which guarantees a fast rate of convergence. Based on a geometric approach to the underlying estimation problem, the CIG method derives its name from the introduction of a new information metric on the mixture parameter space, which is called the component-wise information metric, and serves as a substitute for the Fisher information metric.
Fichier principal
Vignette du fichier
JL_SPL.pdf (2.4 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04270504 , version 1 (04-11-2023)

Identifiants

  • HAL Id : hal-04270504 , version 1

Citer

Jialun Zhou, Salem Said, Yannick Berthoumieu. Adaptive online estimation for mixtures of ECD: a geometric approach. 2023. ⟨hal-04270504⟩
49 Consultations
28 Téléchargements

Partager

Gmail Facebook X LinkedIn More