Mattia Desana

Mattia Desana, Christoph Schnörr

This paper introduces a new probabilistic architecture called Sum-Product Graphical Model (SPGM). SPGMs combine traits from Sum-Product Networks (SPNs) and Graphical Models (GMs): Like SPNs, SPGMs always enable tractable inference using a class of models that incorporate context specific independence. Like GMs, SPGMs provide a high-level model interpretation in terms of conditional independence assumptions and corresponding factorizations. Thus, the new architecture represents a class of probability distributions that combines, for the first time, the semantics of graphical models with the evaluation efficiency of SPNs. We also propose a novel algorithm for learning both the structure and the parameters of SPGMs. A comparative empirical evaluation demonstrates competitive performances of our approach in density estimation.

Mattia Desana, Christoph Schnörr

Sum-Product Networks with complex probability distribution at the leaves have been shown to be powerful tractable-inference probabilistic models. However, while learning the internal parameters has been amply studied, learning complex leaf distribution is an open problem with only few results available in special cases. In this paper we derive an efficient method to learn a very large class of leaf distributions with Expectation-Maximization. The EM updates have the form of simple weighted maximum likelihood problems, allowing to use any distribution that can be learned with maximum likelihood, even approximately. The algorithm has cost linear in the model size and converges even if only partial optimizations are performed. We demonstrate this approach with experiments on twenty real-life datasets for density estimation, using tree graphical models as leaves. Our model outperforms state-of-the-art methods for parameter learning despite using SPNs with much fewer parameters.