Tripartite: Tackling Realistic Noisy Labels with More Precise Partitions.

Area of Science:

• Machine Learning
• Computer Science
• Artificial Intelligence

Background:

• Deep learning models are susceptible to overfitting mislabeled data in large-scale datasets.
• Existing methods often misclassify uncertain noisy samples as clean due to low loss values, degrading model performance.
• Addressing noisy labels is crucial for enhancing the reliability of deep learning models.

Purpose of the Study:

• To propose a novel method, Tripartite, for partitioning training data into uncertain, clean, and noisy subsets.
• To improve the quality of the clean subset by accurately identifying uncertain noisy samples.
• To enhance deep model performance by leveraging clean samples and mitigating the impact of noisy labels.

Main Methods:

• Developed a Tripartite data partitioning strategy based on prediction inconsistencies between two networks and given labels.
• Classified data into three subsets: uncertain, clean, and noisy.
• Applied low-weight learning to uncertain samples and semi-supervised learning to noisy samples.

Main Results:

• Tripartite significantly improves the purity of the clean data subset.
• The proposed method effectively filters out noisy samples with greater precision.
• Experimental results show superior performance compared to state-of-the-art methods on benchmark and real-world datasets.

Conclusions:

• Tripartite offers a more precise approach to handling noisy labels in deep learning.
• The method enhances model performance by better utilizing clean data and reducing the negative effects of mislabeled samples.
• Tripartite demonstrates strong potential for improving the robustness of deep learning models on large-scale, real-world datasets.