Large Deviation Methods for Approximate Probabilistic Inference

Michael Kearns, Lawrence Saul

We study two-layer belief networks of binary random variables in which the conditional probabilities Pr[childlparents] depend monotonically on weighted sums of the parents. In larg…