High-level Rapid Prototyping of Graphical Models

Graphical models provide a flexible framework for capturing probabilistic relationships between large numbers of variables. Because of this flexibility, researchers usually compare a number of possible model topologies and parameter settings for their particular application. However, typically a choice exists between libraries in lower-level languages with faster inference, and libraries in higher-level languages with slower inference. Here we present a rapid-prototyping library for graphical models which combines flexible representation in a high-level language with dynamic compilation to low-level code for performing actual inference. We demonstrate the effectiveness of the resulting library on a set of benchmark images, comparing with a reference implementation in the C++ language. We show that the library facilitates the creation of domain-specific languages for quickly defining and testing graphical models.