Combating Mode Collapse via Offline Manifold Entropy Estimation

Generative Adversarial Networks (GANs) have shown compelling results in various tasks and applications recent years. However, mode collapse remains a critical problem GANs. In this paper, we propose novel training pipeline to address the issue of Different from existing methods, generalize discriminator as feature embedding maximize entropy distributions space learned by discriminator. Specifically, two regularization terms, i.e., Deep Local Linear Embedding (DLLE) Isometric Mapping (DIsoMap), are introduced encourage learn structural information embedded data, such that can be well-formed. Based on well-learned supported discriminator, non-parametric estimator is designed efficiently vectors, playing an approximation maximizing generated distribution. By improving distance most similar samples space, our effectively reduces without sacrificing quality samples. Extensive experimental show effectiveness method which outperforms GAN baseline, MaF-GAN CelebA (9.13 vs. 12.43 FID) surpasses state-of-the-art energy-based model ANIMEFACE dataset (2.80 2.26 Inception score).