MR-GAN: Manifold Regularized Generative Adversarial Networks for Scientific Data

Despite the growing interest in applying generative adversarial networks (GANs) in complex scientific applications, training GANs on scientific data remains a challenging problem from both theoretical and practical standpoints. One reason for this is that the generator is unable to accurately capture the underlying complex manifold structure of the real scientific data using only gradients from the discriminator. In this paper, we address this challenge using a novel approach that exploits the unique geometry of the scientific data to improve the quality of the generated data. Specifically, we improve the training of the GAN using an additional term referred to as a manifold regularizer which encourages the generator to respect the unique geometry of the scientific data manifold and generate high quality data. We theoretically prove that the addition of this regularization term leads to improved performance for different classes of GANs including deep convolutional GAN and Wasserstein GAN. Finally, we carry out performance comparisons on diverse datasets: synthetic data (Gaussian mixture), natural image data (celebrity face images (CelebA)), and scientific experimental data (scanning electron microscopy images of organic crystalline materials). In most of these applications, we find that the proposed manifold regularization-based approach helps in avoiding mode collapse, produces stable training, and leads to significant gains in terms of geometry score compared to its unregularized counterparts.