Mode Regularized Generative Adversarial Networks

• DOI: 10.48550/arxiv.1612.02136
• Publication Date: 2016-01-01
• Authors: Tong Che; Yanran Li; Athul Paul Jacob; Yoshua Bengio; Wenjie Li
• Keywords: Mode (computer interface); Generative model; Generative adversarial network

Although Generative Adversarial Networks achieve state-of-the-art results on a variety of generative tasks, they are regarded as highly unstable and prone to miss modes. We argue that these bad behaviors GANs due the very particular functional shape trained discriminators in high dimensional spaces, which can easily make training stuck or push probability mass wrong direction, towards higher concentration than data generating distribution. introduce several ways regularizing objective, dramatically stabilize GAN models. also show our regularizers help fair distribution across modes distribution, during early phases thus providing unified solution missing problem.