Fast Learning by Bounding Likelihoods in Sigmoid Type Belief Networks