Diagnosing data samples for robust GAN training

Abstract

In the real world, the dataset used to train generative models may include noisy samples. This hinders the model from learning the desired distribution. We propose a method to make robust Generative Adversarial Networks (GANs). We empirically analyze the unstable behavior of GAN training on hard-to-be-learned data and develop efficient measures Log Density Ratio Variance (LDRV) and Log Density Ratio Difference (LDRD) to identify these samples. Using the property that noisy data is hard-to-be-learned, LDRV and LDRD can be used to filter out the noisy data in the training dataset. Furthermore, for clean datasets, forcing the model to focus on hard-to-be-learned data by sample re-weighting using LDRV and LDRD can enhance training performance. Finally, we can make a robust GAN by filtering noisy data and re-weighting.