Multivariate-Information Adversarial Ensemble for Scalable Joint Distribution Matching

摘要: A broad range of cross-$m$-domain generation researches boil down to matching a joint distribution by deep generative models (DGMs). Hitherto algorithms excel in pairwise domains while as $m$ increases, remain struggling scale themselves fit distribution. In this paper, we propose domain-scalable DGM, i.e., MMI-ALI for $m$-domain matching. As an ensemble model ALIs \cite{dumoulin2016adversarially}, is adversarially trained with maximizing Multivariate Mutual Information (MMI) w.r.t. variables each pair and their shared feature. The negative MMIs are upper bounded series feasible losses that provably lead distributions. linearly scales increases thus, strikes right balance between efficacy scalability. We evaluate diverse challenging scenarios verify its superiority.