Improving on Expectation Propagation