Improving FM-GAN Through Mixup Manifold Regularization

Due to their ability to model the manifold of natural images, generative adversarial networks (GANs) have been adapted to semi-supervised learning and shown promising results. Using this property of GANs, manifold regularization has been incorporated into feature-matching GANs (FM-GAN) [14] and shown improvement over FM-GAN. In this work, we present a novel approach for performing manifold regularization in semi- supervised classification using the FM-GAN framework. Our approach, mixup manifold regularization (MUMR), regularizes the networks to favor linear behavior so that linear interpolations of noise vectors lead to linear interpolations of the associated predictions. We show that our approach is able to achieve state-of-the-art results when compared to other GANs- and perturbation-based approaches.