Intervention Adversarial Auto-Encoder

Yang Hu, Cheng Zhang

Keywords: Deep Learning, Generative Models, Adversarial Training

Abstract: In this paper we propose a new method to stabilize the training process of the latent variables of adversarial auto-encoders, which we name Intervention Adversarial auto-encoder (IVAAE). The main idea is to introduce a sequence of distributions that bridge the distribution of the learned latent variable and its prior distribution. We theoretically and heuristically demonstrate that such bridge-like distributions, realized by a multi-output discriminator, have an effect on guiding the initial latent distribution towards the target one and hence stabilizing the training process. Several different types of the bridge distributions are proposed. We also apply a novel use of Stein variational gradient descent (SVGD), by which point assemble develops in a smooth and gradual fashion. We conduct experiments on multiple real-world datasets. It shows that IVAAE enjoys a more stable training process and achieves a better generating performance compared to the vanilla Adversarial auto-encoder (AAE)

One-sentence Summary: We propose a generative model which possess robust and stable training property by leveraging interventions on latent variables and classifying them.