Enhancing Robustness in Learning with Noisy Labels: An Asymmetric Co-Training Approach

Mengmeng Sheng1, Zeren Sun1, Gensheng Pei1, Tao Chen1, Haonan Luo2, Yazhou Yao1
1 Nanjing University of Science and Technology, 2 Southwest Jiaotong University
ACM International Conference on Multimedia (ACM MM), 2024
Paper Code arXiv
Poster Dataset Distillation (PoDD)

We propose a novel asymmetric co-training (ACT) approach to mitigate the negative impact induced by noisy labels. It trains two networks asymmetrically to improve the reliability of learned knowledge. Through this asymmetric training framework, our RTM and NTM can provide more distinctive insights for clean sample selection compared to existing SCT methods. We introduce two novel criteria to establish an asymmetric sample selection and mining strategy based on the relationship between model predictions, focusing on their consensus and disagreement with given labels. Moreover, we propose a dynamic sample re-weighting method, utilizing historical training states to enhance the reliability of our clean sample selection and mining.

Abstract

Label noise, an inevitable issue in various real-world datasets, tends to impair the performance of deep neural networks. A large body of literature focuses on symmetric co-training, aiming to enhance model robustness by exploiting interactions between models with distinct capabilities. However, the symmetric training processes employed in existing methods often culminate in model consensus, diminishing their efficacy in handling noisy labels. To this end, we propose an Asymmetric Co-Training (ACT) method to mitigate the detrimental effects of label noise. Specifically, we introduce an asymmetric training framework in which one model (i.e., RTM) is robustly trained with a selected subset of clean samples while the other (i.e., NTM) is conventionally trained using the entire training set. We propose two novel criteria based on agreement and discrepancy between models, establishing asymmetric sample selection and mining. Moreover, a metric, derived from the divergence between models, is devised to quantify label memorization, guiding our method in determining the optimal stopping point for sample mining. Finally, we propose to dynamically re-weight identified clean samples according to their reliability inferred from historical information. We additionally employ consistency regularization to achieve further performance improvement. Extensive experimental results on synthetic and real-world datasets demonstrate the effectiveness and superiority of our method.

Differences between Symmetric and Asymmetric Co-training Approachs

Self-Adaptive and Class-Balanced Sample Selection and Re-weighting

SCT methods usually entail the simultaneous training of two networks with identical architectures but distinct weight initializatio. The twin networks adopt the same training strategy, capitalizing on their distinct learning capabilities to provide mutual guidance throughout the learning process. In our ACT approach, two models with identical architectures are simultaneously trained utilizing distinct training strategies.

Average Test Accuracy on CIFAR100N and CIFAR80N Under Various Noise Conditions

Average Test Accuracy on CIFAR100N and CIFAR80N Under Various Noise Conditions

Experiments are conducted under symmetric and asymmetric label noise conditions. Results of existing methods are mainly drawn from previous works. † indicates methods re-implemented using their open-sourced code and default hyper-parameters.

Comparison of Test Accuracy on Real-World Noisy Datasets

Comparison of Test Accuracy on Real-World Noisy Datasets

The table shows the experimental results of existing methods and ACT on Web-Aircraft, Web-Bird, and Web-Car datasets. Results of existing methods are mainly drawn from previous works. † indicates methods re-implemented using their open-sourced code and default hyper-parameters.

Comparison of Test Accuracy on Real-World Noisy Datasets

The table shows the experimental results of existing methods and ACT on Food-101N.

BibTeX

@article{sheng2024foster,
  title={Enhancing Robustness in Learning with Noisy Labels: An Asymmetric Co-Training Approach},
  author={Sheng, Mengmeng and Sun, Zeren and Gensheng, Pei and Chen, Tao and Haonan, Luo and Yao, Yazhou},
  journal={ACM International Conference on Multimedia (ACM MM)},
  year={2024}
}