Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Foster Adaptivity and Balance in Learning with Noisy Labels

Mengmeng Sheng, Zeren Sun, Tao Chen, Shuchao Pang, Yucheng Wang, Yazhou Yao

TL;DR

The paper tackles learning with noisy labels by proposing SED, a self-adaptive and class-balanced framework that eliminates reliance on dataset-specific priors. It integrates three components: adaptive sample selection with global and class-wise thresholds, mean-teacher-based label correction for noisy samples, and a dynamic, class-aware re-weighting scheme coupled with consistency regularization to improve generalization. Through extensive experiments on synthetic (CIFAR100N, CIFAR80N) and real-world (Web-Aircraft, Web-Bird, Web-Car) datasets, SED achieves superior or competitive performance and robust ablations confirm the contribution of each module. The approach is practical, scalable, and implemented with standard training protocols, offering improved robustness to both closed-set and open-set label noise without requiring prior threshold tuning or Mixup tricks.

Abstract

Label noise is ubiquitous in real-world scenarios, posing a practical challenge to supervised models due to its effect in hurting the generalization performance of deep neural networks. Existing methods primarily employ the sample selection paradigm and usually rely on dataset-dependent prior knowledge (\eg, a pre-defined threshold) to cope with label noise, inevitably degrading the adaptivity. Moreover, existing methods tend to neglect the class balance in selecting samples, leading to biased model performance. To this end, we propose a simple yet effective approach named \textbf{SED} to deal with label noise in a \textbf{S}elf-adaptiv\textbf{E} and class-balance\textbf{D} manner. Specifically, we first design a novel sample selection strategy to empower self-adaptivity and class balance when identifying clean and noisy data. A mean-teacher model is then employed to correct labels of noisy samples. Subsequently, we propose a self-adaptive and class-balanced sample re-weighting mechanism to assign different weights to detected noisy samples. Finally, we additionally employ consistency regularization on selected clean samples to improve model generalization performance. Extensive experimental results on synthetic and real-world datasets demonstrate the effectiveness and superiority of our proposed method. The source code has been made available at https://github.com/NUST-Machine-Intelligence-Laboratory/SED.

Foster Adaptivity and Balance in Learning with Noisy Labels

TL;DR

The paper tackles learning with noisy labels by proposing SED, a self-adaptive and class-balanced framework that eliminates reliance on dataset-specific priors. It integrates three components: adaptive sample selection with global and class-wise thresholds, mean-teacher-based label correction for noisy samples, and a dynamic, class-aware re-weighting scheme coupled with consistency regularization to improve generalization. Through extensive experiments on synthetic (CIFAR100N, CIFAR80N) and real-world (Web-Aircraft, Web-Bird, Web-Car) datasets, SED achieves superior or competitive performance and robust ablations confirm the contribution of each module. The approach is practical, scalable, and implemented with standard training protocols, offering improved robustness to both closed-set and open-set label noise without requiring prior threshold tuning or Mixup tricks.

Abstract

Label noise is ubiquitous in real-world scenarios, posing a practical challenge to supervised models due to its effect in hurting the generalization performance of deep neural networks. Existing methods primarily employ the sample selection paradigm and usually rely on dataset-dependent prior knowledge (\eg, a pre-defined threshold) to cope with label noise, inevitably degrading the adaptivity. Moreover, existing methods tend to neglect the class balance in selecting samples, leading to biased model performance. To this end, we propose a simple yet effective approach named \textbf{SED} to deal with label noise in a \textbf{S}elf-adaptiv\textbf{E} and class-balance\textbf{D} manner. Specifically, we first design a novel sample selection strategy to empower self-adaptivity and class balance when identifying clean and noisy data. A mean-teacher model is then employed to correct labels of noisy samples. Subsequently, we propose a self-adaptive and class-balanced sample re-weighting mechanism to assign different weights to detected noisy samples. Finally, we additionally employ consistency regularization on selected clean samples to improve model generalization performance. Extensive experimental results on synthetic and real-world datasets demonstrate the effectiveness and superiority of our proposed method. The source code has been made available at https://github.com/NUST-Machine-Intelligence-Laboratory/SED.
Paper Structure (13 sections, 16 equations, 3 figures, 4 tables, 1 algorithm)

This paper contains 13 sections, 16 equations, 3 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Problem Statement
  5. Adaptive and Balanced Sample Selection
  6. Adaptive and Balanced Re-weighting
  7. Overall Framework
  8. Experiments
  9. Experiment Setup
  10. Evaluation on Synthetic Datasets
  11. Evaluation on Real-world Datasets
  12. Ablation Studies
  13. Conclusion

Figures (3)

  • Figure 1: (a-b) Self-adaptive and class-balanced sample selection based on predicted probability w.r.t. given labels. The blue curve indicates the class-specific selection thresholds. (c-d) Self-adaptive and class-balanced sample re-weighting based on correction confidence. The orange curve represents the class-specific confidence threshold.
  • Figure 2: The overall framework of our SED. We first divide the training set into a clean subset and a noisy subset based on global and local thresholds that are dynamically updated. Our threshold design enables self-adaptivity and class balance in sample selection. We then employ a mean-teacher model to correct labels for noisy samples. Based on the correction confidence, SED adaptively assigns different weights to label-corrected noisy samples and uses them for training. Finally, SED further boosts the model performance by imposing an additional consistency regularization loss on selected clean samples. The final objective loss integrates the classification losses on clean and noisy samples and the regularization loss on clean samples.
  • Figure 3: Comparison of different sample selection methods and the ablation results of the parameter $\lambda_m$. (a) The overall precision of sample selection (%) vs. epochs. (b) The class-wise precision of sample selection (%) vs. classes. (c) The test accuracy (%) vs. epochs. (d) The test accuracy (%) of using different $\lambda_m$.