A Robust Deep Learning Framework for Mitigating Label Noise With Dual Selective Attention
The performance of Machine Learning (ML) models is highly sensitive to data quality, still the impact of label accuracy remains underexplored. In this study, a novel architecture, the Dual Selective Attention Network (DSAN), is proposed to improve robustness against mislabeled data in deep learning...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11062817/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The performance of Machine Learning (ML) models is highly sensitive to data quality, still the impact of label accuracy remains underexplored. In this study, a novel architecture, the Dual Selective Attention Network (DSAN), is proposed to improve robustness against mislabeled data in deep learning tasks. DSAN incorporates Position Attention Module (PAM) and Channel Attention Module (CAM) to emphasize relevant spatial and channel level features, effectively suppressing the influence of incorrect labels. DSAN was evaluated alongside four baseline models Visual Geometry Group Network (VGG16), Convolutional Neural Network (CNN), Artificial Neural Network (ANN), and ResNet-50 on three datasets, with label noise introduced at 0%, 5%, 10%, 15%, and 20% to simulate real-world mislabeling scenarios. Model performance was assessed using accuracy, precision, recall (sensitivity), and F1-score. DSAN maintained over 96.5% accuracy under 20% label noise and outperformed all baseline models by 4–8% on average across performance metrics. While CNN showed moderate robustness and ResNet-50 exhibited better resilience due to its residual learning mechanism, both models experienced notable performance degradation as noise increased. VGG16 and ANN were particularly vulnerable, with sharp declines observed even under low noise levels. To further address mislabeling, the Latent Space Variation using Supervised Autoencoder (AQUAVS) technique was also applied to remove mislabeled data and assess model recovery. Although slight improvements were observed, AQUAVS still lagged behind DSAN in all scenarios. Additionally, an ablation study was conducted to evaluate the individual contributions of PAM and CAM, showing that their combination significantly enhances DSAN robustness to mislabeled data while maintaining high classification performance under label noise. These results highlight the importance of designing architectures that are inherently robust to label noise. By maintaining high reliability under noisy conditions, DSAN contributes to the development of more dependable and generalizable AI systems. |
|---|---|
| ISSN: | 2169-3536 |