Belief propagation is a universal method used in many field, such as AI, statistical physical, and error correction codes. It gives the exact inference when a graph is tree, but also a good approximation even if it is loopy. The authors have developed an information geometrical framework to analyze the belief propagation algorithm, which gives a unified view. In this article, the authors show the idea of the belief propagation algorithm, the information geometrical framework, and some results of the analysis.
抄録全体を表示