Spatial Triplet Markov Trees for auxiliary variational inference in Spatial Bayes Networks

Hugo Gangloff (1, 2) , Jean-Baptiste Courbot (3) , Emmanuel Monfrini (4, 5, 6) , Christophe Collet (1)

1 ICube - Laboratoire des sciences de l'ingénieur, de l'informatique et de l'imagerie
2 Groupe Européen de Recherche sur les Prothèses appliquées à la Chirurgie Vasculaire, Strasbourg
3 IRIMAS - Institut de Recherche en Informatique Mathématiques Automatique Signal
4 IP Paris - Institut Polytechnique de Paris
5 TSP - CITI - Communications, Images et Traitement de l'Information
6 TIPIC-SAMOVAR - Traitement de l'Information Pour Images et Communications

In this article, we develop a Triplet Markov Tree model with auxiliary random variables for an approximate inference in an intractable probabilistic model. It is based on recent advances on probabilistic modeling and varational inference with auxiliary random variables. The new Triplet Markov Tree model performs better than the classical Mean-Field variational inference and than a tree-structured variational inference. Our study provides insights and motivations for the developing workaround models involving auxiliary random variables.

Mots clés: Variational inference; Auxiliary random variables; Triplet Markov models

Domaines: Méthodologie [stat.ME]

Hugo Gangloff, Jean-Baptiste Courbot, Emmanuel Monfrini, Christophe Collet. Spatial Triplet Markov Trees for auxiliary variational inference in Spatial Bayes Networks. SMTDA 2020: 6th international conference on Stochastic Modeling Techniques and Data Analysis, Jun 2020, Barcelone (online), Spain. pp.237-249. ⟨hal-03122810⟩