Bayesian model selection for deep exponential families
I. A. Grenier · eScholarship@McGill (McGill) · 2016
In their article "Deep Exponential Families" , Ranganath, Tang, Charlin and Blei (2014) introduce deep exponential families (DEFs), a special type of hierarchical models under which the layers of latent variables are linked through their canonical parameters. The goal of this thesis is to provide an ecient model selection technique for DEFs. The focus has been set on multinomial-like datasets generated from a Poisson DEFs which are analogous to classication problems. By using Markov Chain Monte Carlo sampling, we are able to look at Bayesian and frequentists predictive measures to achieve model selection. Finally, to assess the need for more complex systems, counts generated from a mixture of multinomials are studied under both the regular topic model approach and the DEFs modeling.