Beliefs propagation in log domain: a neural inspired algorithm for machine learning

In this paper, we consider a variant of belief propagation algorithm in a tree graphical model where computations are carried out in the negative log-likelihood domain. Unlike the min-product algorithm, our goal is not limited to estimating the mode of the marginal distribution. We would like to obtain the entire marginal distribution as the sum-product algorithm does. We applied the algorithm to learn effective users features for A/B testing. We discussed scalable extension to the proposed algorithm for processing large amount of data.The primary goal of a parallel program is to reduce running time comparing to the sequential program by taking full advantage of computing power of multiprocessors.