Graphical Generative Adversarial Networks – Arxiv Vanity

We first assume that the recognition model can also be factorized in the same way. Then we're going to minimize the divergence in terms of each factor individually. Though the approximation cannot be justified theoretically, some intuition and empirical results [Minka(2005)] suggest that the gap is small if the approximate posterior is a good fit to the true posterior. In our approach, we make the same assumption because q(A) will be cancelled in the likelihood ratio if D belongs to f-divergence and we can ignore other factors when discriminating q(A), which reduces the complexity of the problem. See Appendix B for the derivation.