Reviews: AdaGAN: Boosting Generative Models

AdaGAN is a meta-algorithm proposed for GAN. The key idea of AdaGAN is: at each step reweight the samples and fits a generative model on the reweighted samples. The final model is a weighted addition of the learned generative models. The main motivation is to reduce the mode-missing problem of GAN by reweighting samples at each step. It is claimed in the Introduction that AdaGAN can also use WGAN or mode-regularized GAN as base generators (line 55).