Tackling mode collapse in multi-generator GANs with orthogonal vectors

Abstract

Generative Adversarial Networks (GANs) have been widely used to generate realistic-looking instances. However, training robust GAN is a non-trivial task due to the problem of mode collapse. Although many GAN variants are proposed to overcome this problem, they have limitations. Those existing studies either generate identical instances or result in negative gradients during training. In this paper, we propose a new approach to training GAN to overcome mode collapse by employing a set of generators, an encoder and a discriminator. A new minimax formula is proposed to simultaneously train all components in a similar spirit to vanilla GAN. The orthogonal vector strategy is employed to guide multiple generators to learn different information in a complementary manner. In this way, we term our approach Multi-Generator Orthogonal GAN (MGO-GAN). Specifically, the synthetic data produced by those generators are fed into the encoder to obtain feature vectors. The orthogonal value is calculated between any two feature vectors, which loyally reflects the correlation between vectors. Such a correlation indicates how different information has been learnt by generators. The lower the orthogonal value is, the more different information the generators learn. We minimize the orthogonal value along with minimizing the generator loss through back-propagation in the training of GAN. The orthogonal value is integrated with the original generator loss to jointly update the corresponding generator’s parameters. We conduct extensive experiments utilizing MNIST, CIFAR10 and CelebA datasets to demonstrate the significant performance improvement of MGO-GAN in terms of generated data quality and diversity at different resolutions.