Dual Discriminator Generative Adversarial Nets (1709.03831v1)

Abstract: 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.

Analysis of "Dual Discriminator Generative Adversarial Nets"

The paper "Dual Discriminator Generative Adversarial Nets" by Tu Dinh Nguyen, Trung Le, Hung Vu, and Dinh Phung introduces a novel approach to address the notable issue of mode collapse in Generative Adversarial Networks (GANs). This problem results in a lack of diversity in generated samples, where the generator produces samples that only cover a limited portion of the data distribution.

Contributions

The authors propose Dual Discriminator Generative Adversarial Networks (D2GAN), which incorporates two discriminators instead of the singular discriminator utilized in traditional GANs. The theoretical foundation lies in combining the Kullback-Leibler (KL) and the reverse KL divergences into a unified objective function, thereby exploiting their complementary properties. By balancing these divergences, D2GAN effectively reduces the issues of mode collapse that GAN systems often encounter.

Theoretical Framework

The D2GAN model is structured as a three-player minimax game involving two different discriminators and one generator:

• Discriminator $D_1$ rewards samples that belong to the true data distribution.
• Discriminator $D_2$ favors samples from the generator.
• Generator $G$ aims to deceive both discriminators by generating plausible samples.

The method manipulates the GAN's adversarial setup to minimize both the KL divergence from data to model distribution and the reverse KL divergence. This duality allows the generator to learn a balanced strategy that neither overgeneralizes nor collapses into fewer modes.

Experimental Evaluation

The authors conducted extensive experiments using synthetic and real-world datasets, such as MNIST, CIFAR-10, STL-10, and the large-scale ImageNet dataset. The experiments were designed to compare D2GAN with other GAN variants based on both qualitative and quantitative metrics, utilizing performance measures such as the Inception Score and MODE Score.

Key Findings:

• D2GAN demonstrates superior performance over traditional GANs and several state-of-the-art GAN variants in terms of generating diverse and high-quality samples.
• On benchmark datasets like MNIST-1K—a synthetic dataset with increased complexity—D2GAN effectively captured all modes of the data distribution. It showed improved reverse KL scores, indicative of better mode coverage compared to competitors like UnrolledGAN and Reg-GAN.
• On real-world datasets, D2GAN achieved competitive Inception Scores, affirming its scalability and effectiveness in generating diverse samples on larger and more challenging image datasets.

Implications

The significant improvement in generating diverse and high-quality data samples suggests that integrating dual discriminators can provide a robust mechanism against mode collapse in GANs. Importantly, the framework's ability to scale effectively to large datasets such as ImageNet demonstrates its viability in practical applications, potentially enhancing tasks ranging from image synthesis to data augmentation in machine learning pipelines.

Future Directions

While D2GAN provides a substantial improvement over existing GAN frameworks, there is room for further enhancement by integrating other advanced techniques. For example, combining D2GAN with autoencoders or adopting conditional GAN strategies could further enhance sample quality and diversity. Additionally, exploring the balance between the two divergence measures, particularly in the face of varying dataset complexities, could uncover nuanced insights into fine-tuning GAN architectures.

In conclusion, Dual Discriminator Generative Adversarial Nets present a promising advancement in the field of generative modeling, particularly in tackling the notorious challenge of mode collapse in GANs. The implementation of a dual discriminator framework not only offers theoretical improvements but also delivers practical enhancements in generating diverse and realistic data samples.