A variational approximation for Bayesian networks with discrete and continuous latent variables

摘要: We show how to use a variational approximation the logistic function perform approximate inference in Bayesian networks containing discrete nodes with continuous parents. Essentially, we convert Gaussian, which facilitates exact inference, and then iteratively adjust parameters improve quality of approximation. demonstrate experimentally that this is much faster than sampling, but comparable accuracy. also introduce simple new technique for handling evidence, allows us handle arbitrary distributionson observed nodes, as well achieving significant speedup variables large cardinality.