Modeling the distribution of natural images is challenging, partly because of strong statistical dependencies which can extend over hundreds of pixels. recurrent neural networks have been successful in capturing long-range dependencies in a number of problems, but only recently have fo