In Bayesian network theory, the joint conditional posterior $P(h|v)$ is often intractable but the joint posterior $P(h,v)$ is tractable. (Where h are hidden nodes and v are evidence nodes). Variational Bayes (VB) seeks to approximate the conditional by some $Q(h|\theta)$ whose form is chosen so it can be optimised. For example, the 'mean field' approximation assumes that Q factors the product of terms in h.
