Gamma Probabilistic Databases: Learning from Exchangeable Query-Answers

Abstract

In this paper we propose a novel knowledge compilation technique that compiles Bayesian inference procedures, starting from probabilistic programs expressed in terms of probabilistic queryanswers. To do so, we extend the framework of Dirichlet Probabilistic Databases with the ability to process exchangeable observations of query-answers. We show that the resulting framework can encode non-trivial models, like Latent Dirichlet Allocation and the Ising model, and generate high-performance Gibbs samplers for both models.