Séance Séminaire

Séminaire de Probabilités et Statistique

Online inference in general hidden Markov models with intractable densities

This talk will consider general hidden Markov models , i.e., a model where observed data is supposed to be a random variable whose distribution depend on an unknown sequence of hidden states, which is supposed to be Markovian. I will consider the case when either the transition density of the latent state or the conditional likelihood of an observation given a state is intractable.
In this setting, obtaining low variance estimators of expectations under the posterior distributions of the unobserved states given the observations is a challenging task.
I will present how, thanks to a recent pseudo marginal sequential Monte Carlo algorithm, we can perform online estimation (such as online recursive maximum likelihood estimation) in such models. The proposed estimator is consistant and asymptotically normal.
I’ll present applications in the context of stochastic neural networks, and to multivariate partially observed diffusion processes (i.e., when the hidden dynamics is solution to a stochastic differential equation).

This talk is based on two recent papers : https://hal.archives-ouvertes.fr/hal-02194237 and https://hal.archives-ouvertes.fr/hal-02476102

Inscription pour assister à l'exposé sur place (20 personnes maximum) : https://framaforms.org/seminaire-de-probabilites-et-statistiques-1622709244

Webinaire ouvert à toutes et tous : https://umontpellier-fr.zoom.us/j/85813807839