Variational Implicit Distribution Matching

Hiroshi Wakimori, Hiromasa Takei, Tikara Hosino

Abstract: Modeling with implicit distributions such as GAN is effective technique in the probabilistic generative models because of its flexibility and few assumptions on the target distribution. For associating two distributions, ALI and other variants were proposed with joint distribution matching. However, they have disadvantages in the balance of the mode missing behavior and sample reconstruction quality.In this paper, we propose variational implicit distribution matching which is natural probabilistic extension of joint matching to its marginal and conditional distributions. We experimentally show that our proposal achieves the balance of the reconstruction and random generation quality and is competitive to the state of the art.

TL;DR: We propose variational implicit distribution matching which is natural probabilistic extension of joint matching to its marginal and conditional distributions.

Keywords: variational method, implicit model, generative model