Despite its popularity, self-training is well-believed to be unreliable and often leads to training instability.

The ability of machine learning to generate structured objects has progressed dramatically in the lastfewyears.

For example, in image object recognition and language models building[4,5,6,7,8,9], the number of classes scales as the number of possible objects or the dictionary size respectively.

Neural Information Processing Systems http://nips.cc/

We study whether and how can we model a joint distributionp(x,z)using two conditional modelsp(x|z) and q(z|x) that form a cycle.

Work done during an internship at Microsoft Research Asia.