Full Conditional Free Energy Based Inference

Inference on graphical models has great applications in the fields such as pattern recognition, artificial intelligence and statistics. The inference problem is usually studied as an optimization problem w.r.t. free energy, such as Bethe/Kikuchi free energy minimization. However, due to the nonconvexity of these free energies, it is often infeasible to obtain the global optimum. In this paper, we propose a new inference approach that can obtain the global optimum. Subsequently, we interpret this approach in terms of minimizing a new free energy, full conditional free energy (FCFE). Based on FCFE, approximate FCFE and an efficient approximate algorithm are proposed. Finally, experiments show the efficiency of the inference framework.