Estimation de mélange de Gaussiennes sur données compressées

Anthony Bourrier (1, 2) , Rémi Gribonval (2) , Patrick Pérez (1)

1 Technicolor [Cesson Sévigné]
2 PANAMA - Parcimonie et Nouveaux Algorithmes pour le Signal et la Modélisation Audio

Estimating a probability mixture model from a set of vectors typically requires a large amount of memory if the data is voluminous. We propose a framework where the data is jointly compressed to a fixed-size representation called sketch, composed of empirical moments calculated from the data. By analogy with compressive sensing, we derive a parameter estimation algorithm from the sketch. We experimentally show that our algorithm allows precise estimation while consuming less memory than an EM algorithm for voluminous data. The algorithm also provides a privacy-preserving estimation tool since the sketch does not disclose information about individual datum it is based on.

Mots clés: Gaussian Mixture Estimation; Compressive Sensing; Database Sketch; Compressive Learning

Domaines: Traitement du signal et de l'image [eess.SP]; Traitement du signal et de l'image [eess.SP]

Anthony Bourrier, Rémi Gribonval, Patrick Pérez. Estimation de mélange de Gaussiennes sur données compressées. 24ème Colloque Gretsi, Sep 2013, France. pp.191. ⟨hal-00839579⟩