GANs with Suppressed Mode Collapse Using Intrinsic Rewards
Toi Tsuneda, Taro Kiriyama, Kosuke Shintani, Satoshi Yamane · 2021
In recent years, generative adversarial net-works (GANs) have attracted much attention because it can generate new data and information by taking advantage of the characteristics of samples. However, there are many problems with GANs, such as difficulty in learning well and instability. In this paper, we proposed a method to suppress mode collapse, which is one of the problems that often occur in GANs, by utilizing intrinsic rewards used in deep reinforcement learning. We have conducted comparative experiments between the pro-posed method and the original GANs in several ways to show the performance of the proposed method.