Learning as MAP Inference in Discrete Graphical Models