Conditional GANs with Auxiliary Discriminative Classifier (2107.10060v5)

Abstract: Conditional generative models aim to learn the underlying joint distribution of data and labels to achieve conditional data generation. Among them, the auxiliary classifier generative adversarial network (AC-GAN) has been widely used, but suffers from the problem of low intra-class diversity of the generated samples. The fundamental reason pointed out in this paper is that the classifier of AC-GAN is generator-agnostic, which therefore cannot provide informative guidance for the generator to approach the joint distribution, resulting in a minimization of the conditional entropy that decreases the intra-class diversity. Motivated by this understanding, we propose a novel conditional GAN with an auxiliary discriminative classifier (ADC-GAN) to resolve the above problem. Specifically, the proposed auxiliary discriminative classifier becomes generator-aware by recognizing the class-labels of the real data and the generated data discriminatively. Our theoretical analysis reveals that the generator can faithfully learn the joint distribution even without the original discriminator, making the proposed ADC-GAN robust to the value of the coefficient hyperparameter and the selection of the GAN loss, and stable during training. Extensive experimental results on synthetic and real-world datasets demonstrate the superiority of ADC-GAN in conditional generative modeling compared to state-of-the-art classifier-based and projection-based conditional GANs.

Conditional GANs with Auxiliary Discriminative Classifier

The paper presents a novel approach to enhance Conditional Generative Adversarial Networks (cGANs) with a proposed model termed ADC-GAN, which includes an Auxiliary Discriminative Classifier. This advancement addresses key limitations observed in traditional Auxiliary Classifier GANs (AC-GANs), specifically targeting the issue of low intra-class diversity within generated samples.

Key Insights and Methodology

The authors identify that the root of the low intra-class diversity problem in AC-GANs lies in the generator-agnostic nature of the classifier, which fails to provide adequate guidance for the generator to accurately learn the joint distribution of data and labels. ADC-GAN resolves this by making the classifier generator-aware, allowing it to distinguish both real and generated data while simultaneously recognizing their class labels. This is achieved via a discriminative classifier that operates similarly to the discriminator but is tasked with classification between real and synthetic data.

The theoretical underpinning of ADC-GAN demonstrates that it can robustly learn the joint distribution without the original discriminator's assistance. The method is resilient to variations in the hyperparameter settings and the choice of GAN loss functions, ensuring stable training.

Empirical Validation

Experiments conducted on both synthetic and real-world datasets, such as CIFAR-10, CIFAR-100, and Tiny-ImageNet, show that ADC-GAN outperforms existing classifier-based and projection-based cGANs. Notably, ADC-GAN achieves superior Fréchet Inception Distance (FID) and Intra-FID scores, indicating enhanced overall and intra-class image quality. The model also demonstrates improved handling of training stability and data-to-class relations, verified through t-SNE visualizations and classification accuracy assessments.

Comparisons with Other Methods

The paper provides a detailed comparison of ADC-GAN with TAC-GAN, PD-GAN, and others, elucidating potential issues inherent in these approaches. TAC-GAN attempts to mitigate AC-GAN's diversity problems with twin classifiers but faces training instability. PD-GAN, relying on projection discriminators, lacks partition terms to model complete data-label dependencies, thereby affecting fidelity.

Implications and Future Directions

ADC-GAN's ability to accurately model the data-label distribution and improve intra-class diversity can significantly bolster the practical applications of cGANs in fields that require high-quality conditional image generation. Future research may explore extending ADC-GAN's principles to more complex datasets and tasks, as well as evaluating its integration with increasingly sophisticated discriminative structures.

In summary, this work proposes a methodologically sound and empirically validated approach to enhance conditional generative modeling, providing a valuable contribution to the development and refinement of generative models.