Dual Discriminator Generative Adversarial Nets

By Tu Dinh Nguyen, Trung Le, Hung Thanh Vu, Dinh Q. Phung. 2017

346 citations recorded by the source corpus.

We propose in this paper a novel approach to tackle the problem of mode collapse encountered in generative adversarial network (GAN). Our idea is intuitive but proven to be very effective, especially in addressing some key limitations of GAN. In essence, it combines the Kullback-Leibler (KL) and reverse KL divergences into a unified objective function, thus it exploits the complementary statistical properties from these divergences to effectively diversify the estimated density in capturing multi-modes. We term our method dual discriminator generative adversarial nets (D2GAN) which, unlike GAN, has two discriminators; and together with a generator, it also has the analogy of a minimax game, wherein a discriminator rewards high scores for samples from data distribution whilst another discriminator, conversely, favoring data from the generator, and the generator produces data to fool both two discriminators. We develop theoretical analysis to show that, given the maximal discriminators, optimizing the generator of D2GAN reduces to minimizing both KL and reverse KL divergences between data distribution and the distribution induced from the data generated by the generator, hence effectively avoiding the mode collapsing problem. We conduct extensive experiments on synthetic and real-world large-scale datasets (MNIST, CIFAR-10, STL-10, ImageNet), where we have made our best effort to compare our D2GAN with the latest state-of-the-art GAN's variants in comprehensive qualitative and quantitative evaluations. The experimental results demonstrate the competitive and superior performance of our approach in generating good quality and diverse samples over baselines, and the capability of our method to scale up to ImageNet database.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

AutoGAN: Neural Architecture Search for Generative Adversarial Networks (2019, neutral): “Method Inception score FID D2GAN [44] 7.”
Mutual Information Neural Estimation (2018, background): “Mode collapse (Che et al., 2016; Dumoulin et al., 2016; Donahue et al., 2016; Salimans et al., 2016; Metz et al., 2017; Saatchi & Wilson, 2017; Nguyen et al., 2017; Lin et al., 2017; Ghosh et al., 2017) is a common pathology of generative adversarial networks (GANs, Goodfellow et al., 2014), where…”
PacGAN: The Power of Two Samples in Generative Adversarial Networks (2017, neutral): “, [16], [18], [19], [20], [21], [22], [23], [24].”
Photographic Text-to-Image Synthesis with a Hierarchically-Nested Adversarial Network (2018, background): “B uses multiple discriminators with one generator [9, 28].”
Generative Adversarial Networks (GANs) (2020, background): “[21] surveyed threat of adversarial attacks on deep learning.”
MGAN: Training Generative Adversarial Nets with Multiple Generators (2018, methodology): “It is worthy to note that only models trained in a completely unsupervised manner without label information are included for fair comparison; and DCGAN’s and D2GAN’s results on STL-10 are available only for the models trained on 32 × 32 resolution.”
Generalized Energy Based Models (2020, background): “While we do not address the mode collapse problem, Xu et al. (2018); Nguyen et al. (2017) showed that KL-based losses are resilient to it thanks to the zero-avoiding property of the KL, a good sign for KALE which is derived from KL by Fenchel duality.”
A Comprehensive Survey of AI-Generated Content (AIGC): A History of Generative AI from GAN to ChatGPT (2023, methodology): “To address mode collapse and improve the generative model, both the D2GAN [80] and GMAN [81] methods extend the traditional GANs by combining extra discriminators.”
Comparing Generative Adversarial Network Techniques for Image Creation and Modification (2018, methodology): “More experiments with methods that improve the stability of GANs are needed, such as spectral normalization [Takeru Miyato, 2018] and using two discriminators [Nguyen et al., 2017].”
Dropout-GAN: Learning from a Dynamic Ensemble of Discriminators (2018, methodology): “, 17], or even increasing both the number of generators and discriminators [14, 10, 5], we turn our focus on methods that solely increases the number of discriminators to prevent mode collapse. D2GAN [29] proposed a single generator dual discriminator architecture where one Drewards samples coming f”
A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications (2020, background): “D2GAN is analogous to a minimax game, wherein one discriminator gives high scores for samples from generated distribution whilst the other discriminator, conversely, favoring data from the true distribution, and the generator generates data to fool both discriminators.”
Self-supervised GAN: Analysis and Improvement with Multi-class Minimax Game (2019, methodology): “Note that, since optimizing KL divergence is not easy because it is asymmetric and could be biased to one direction [32], in our implementation, we use a slightly modified version as described in the Appendix.”
Evolutionary Generative Adversarial Networks (2018, background): “[45] proposed a dual discriminator GAN, which combines the KL and reverse KL divergences into a unified objective function through employing two discriminators.”
IntroVAE: Introspective Variational Autoencoders for Photographic Image Synthesis (2018, methodology): “To synthesize high-resolution images, several studies have trained GANs in a Laplacian pyramid [6, 41] or a tree-like structure [42, 43] with multi-scale discriminators [11, 29, 39], mostly in a coarse-to-fine manner, including the state-of-the-art PGGAN method [18].”
On Data Augmentation for GAN Training (2020, methodology): “Instead of using only one discriminator or generator like standard GAN, the mixture models are proposed, such as multiple discriminators [39], [40], [41], the mixture of generators [42], [43] or an attacker applied as a new player for GAN training [44].”