Abstract
This paper presents the branching approach, an improved deterministic implementation method (such as variational inference and expectation propagation) for Bayesian learning of mixture distributions. The proposed approach uses a set of artifi-cial conditions defined by latent (hidden) variables of the mixture distribution. This condition set is updated iteratively by branching of a condition selected from the previous set. The approximated Bayesian inference is obtained by combining the conditional inferences under all conditions in the set. The proposed approach is compared with several standard implementation methods by using a mixture of normal distributions as an example.