The Linearization of Pairwise Markov Networks.

摘要: Belief Propagation (BP) allows to approximate exact probabilistic inference in graphical models, such as Markov networks (also called random fields, or undirected models). However, no convergence guarantees for BP are known, general. Recent work has proposed by linearizing the update equations around default values special case when all edges network carry same symmetric, doubly stochastic potential. This linearization led guarantees, considerable speed-up, while maintaining high quality results network-based classification (i.e. we only care about most likely label class each node and not probabilities). The present paper generalizes our prior on Linearized (LinBP) with an approach that approximates Loopy any pairwise problem of solving a linear equation system.