Label-Noise Robust Generative Adversarial Networks (1811.11165v2)

Abstract: Generative adversarial networks (GANs) are a framework that learns a generative distribution through adversarial training. Recently, their class-conditional extensions (e.g., conditional GAN (cGAN) and auxiliary classifier GAN (AC-GAN)) have attracted much attention owing to their ability to learn the disentangled representations and to improve the training stability. However, their training requires the availability of large-scale accurate class-labeled data, which are often laborious or impractical to collect in a real-world scenario. To remedy this, we propose a novel family of GANs called label-noise robust GANs (rGANs), which, by incorporating a noise transition model, can learn a clean label conditional generative distribution even when training labels are noisy. In particular, we propose two variants: rAC-GAN, which is a bridging model between AC-GAN and the label-noise robust classification model, and rcGAN, which is an extension of cGAN and solves this problem with no reliance on any classifier. In addition to providing the theoretical background, we demonstrate the effectiveness of our models through extensive experiments using diverse GAN configurations, various noise settings, and multiple evaluation metrics (in which we tested 402 conditions in total). Our code is available at https://github.com/takuhirok/rGAN/.

Label-Noise Robust Generative Adversarial Networks

The paper "Label-Noise Robust Generative Adversarial Networks" presents a novel approach to improve the robustness of class-conditional generative adversarial networks (GANs) in the presence of noisy labels. Given that accurate class-labeled data can be difficult to obtain in real-world scenarios, the proposed solutions, rGANs, incorporate noise transition models with the objective of aligning generative distributions with clean labeled data distributions, even when only noisy labels are accessible during training.

Principal Contributions

The authors introduce two variants of rGANs: rAC-GAN and rcGAN. rAC-GAN bridges auxiliary classifier GAN (AC-GAN) with label-noise robust classification models, leveraging a noise transition model. rcGAN, an extension of conditional GAN (cGAN), tackles label noise without relying on an auxiliary classifier, thus circumventing the diversity issues that arise when generating classifiable images. Both models are engineered to disentangle representations despite noisy input labels, as demonstrated by empirical evaluations.

The paper meticulously constructs a strong theoretical foundation, supporting the assertion that the proposed models narrow the gap between noisy label-conditioned and clean label-conditioned generative distributions. The viability of these models is bolstered through comprehensive experiments involving numerous GAN configurations, diverse noise settings, and multiple evaluation metrics, including the Fréchet Inception Distance (FID) and the GAN-test.

Experimental Evaluation

Across 402 total conditions, the experiments reveal that rGANs consistently deliver superior performance in terms of conditional image generation when compared with baseline models like AC-GAN and cGAN. Specifically, the rcGAN model shows robustness in both symmetric and asymmetric noise scenarios, indicating its potential for real-world applications where label noise is prevalent. This result is further substantiated through extensive analysis of additional metrics such as Intra FID and GAN-train, which demonstrate significant improvements in conditional generative distribution quality and diversity over traditional methods.

The use of estimated noise transition matrices (denoted as $T'$) through a robust two-stage training algorithm exhibits practical applicability in real-world scenarios. The models retain efficacy even when $T'$ slightly diverges from the true noise transition matrix, although the performance degrades at higher noise rates, particularly for CIFAR-100 due to class count and noise complexity.

Advanced Techniques for Noisy Environments

To enhance robustness in severely noisy environments, particularly at a label corruption rate of 90%, the authors introduce mutual information regularization. This technique reinforces the connection between generated images and their corresponding labels, thus mitigating the performance degradation observed in the baseline models under extreme noise conditions.

Implications and Future Work

The research delineates both theoretical and practical implications for the development of noise-resilient conditional generative models. By effectively addressing label noise—an inherently challenging problem in machine learning—the proposed rGANs fundamentally enhance the fidelity and diversity of generated samples. This holds substantial potential for applications in domains where label accuracy cannot be guaranteed, including automated data augmentation, synthetic data generation for training, and enhanced diversity in generative tasks.

Future work could delve into adapting these frameworks for other conditional generative modeling types, such as VAEs or ARs, and exploring further refinements in noise estimation techniques to accommodate even more complex real-world noise distributions. Building models with inherent robustness in representation learning could lead to broader and more reliable applications of GANs in real-world data-intensive tasks.