Fast distributed MAP inference for large-scale graphical models

In every domain of life and society, real-world data gains pull for both a more informed decision making and citizenship. Social and human phenomena carry intricate and unknown dependencies unreachable by traditional machine learning approaches, like regression or classification. How to extract value from large amounts of complex and noisy data? Assuming we know the generative model of our data, inference itself is a combinatorial problem. In this work we put forward a distributed, approximate inference method that attains better accuracy than the centralized LP relaxation of the inference problem, even when the solution of the LP is improved by a local nonconvex method.