Limited Discriminator GAN using explainable AI model for overfitting problem
Jiha Kim, Hyunhee Park
Abstract
Data-driven learning is the most representative deep learning method. Generative adversarial networks (GANs) are designed to generate sufficient data to support such learning. The learning process of GAN models typically trains a generator and discriminator in turn. However, overfitting problems occur when the discriminator depends excessively on the training data. When this problem persists, the image created by the generator shows a similar appearance to the learning image. Images similar to learning images eventually lose the meaning of data augmentation. In this paper, we propose a limited discriminator GAN (LDGAN) model that explains the results of GAN, which is a model that cannot be analyzed externally, such as a black box. The part explained in LDGAN becomes the discriminator model of GAN, and it is possible to check which area of the image is used as the basis for determining fake/real by the discriminator. In the end, a method for limiting the learning of discriminator is proposed based on the described results. Through this, it is possible to avoid the overfitting problem of the discriminator and to generate various images different from the learning image. The LDGAN method allows users to perform meaningful data augmentation with only specific objects except for complex images or backgrounds that require analysis. Compare the LDGAN method with the existing DCGAN and present the extensive simulation results. The extensive simulation result shows that the image generated by the proposed LDGAN including the estimation area is about 10% more.