Aggregation of probabilistic PCA mixtures with a variational-Bayes technique over parameters

Pierrick Bruneau (1, 2) , Marc Gelgon (1, 2) , Fabien Picarougne (2)

1 ATLAS - Complex data management in distributed systems
2 LINA - Laboratoire d'Informatique de Nantes Atlantique

This paper proposes a solution to the problem of aggre- gating versatile probabilistic models, namely mixtures of probabilistic principal component analyzers. These models are a powerful generative form for capturing high-dimensional, non Gaussian, data. They simulta- neously perform mixture adjustment and dimensional- ity reduction. We demonstrate how such models may be advantageously aggregated by accessing mixture pa- rameters only, rather than original data. Aggregation is carried out through Bayesian estimation with a specific prior and an original variational scheme. Experimental results illustrate the effectiveness of the proposal.

Domaines: Informatique [cs]

Pierrick Bruneau, Marc Gelgon, Fabien Picarougne. Aggregation of probabilistic PCA mixtures with a variational-Bayes technique over parameters. IEEE/IAPR Int. Conf. on Pattern Recognition, Aug 2010, Istambul, Turkey. pp.702 - 705. ⟨hal-00471313⟩