An Inference Language for Imaging

We introduce iLang, a language and software framework for probabilistic inference. The iLang framework enables the definition of directed and undirected probabilistic graphical models and the automated synthesis of high performance inference algorithms for imaging applications. The iLang framework is composed of a set of language primitives and of an inference engine based on a message-passing system that integrates cutting-edge computational tools, including proximal algorithms and high performance Hamiltonian Markov Chain Monte Carlo techniques. A set of domain-specific highly optimized GPU-accelerated primitives specializes iLang to the spatial data-structures that arise in imaging applications. We illustrate the framework through a challenging application: spatio-temporal tomographic reconstruction with compressive sensing.