B1615
Title: Stochastic variationally inference for multivariate latent gaussian models
Authors: Gianmarco Callegher - University of Goettingen (Italy) [presenting]
Thomas Kneib - University of Goettingen (Germany)
Paul Wiemann - University of Wisconsin-Madison (United States)
Johannes Soeding - University of Goettingen (Germany)
Abstract: Latent Gaussian models are a subclass of the so-called structured additive models. In structured additive distributional regression, the conditional distribution of the response variables given the covariate information and the vector of model parameters is modelled by means of a P-parametric probability density function where each parameter is modelled through a linear predictor (e.g. linear effects, random effects, Bayesian penalized splines, Gaussian Markov random fields) and a bijective response function that maps the domain of the predictor into the domain of the parameter. A method is presented to perform inference in multivariate latent Gaussian models (mGLMs) using stochastic variational inference, a technique for approximating posterior distributions through optimization. The idea is to define a family of densities over the latent variables defined by a vector of variational parameters and then find the settings of the parameters that make the variational distribution close to the posterior by stochastic optimization.