Learning With Noisy Labels via Self-Reweighting From Class Centroids.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) require large labeled datasets for effective training.
• Noisy labels in training data can significantly degrade DNN performance.
• Existing methods for handling noisy labels using dynamic sample weighting are often unreliable.

Purpose of the Study:

• To develop a robust method for training DNNs with noisy labels.
• To improve the discriminative ability of DNNs when trained on corrupted data.
• To enhance the robustness of decision boundaries against label noise.

Main Methods:

• Proposes Self-Reweighting from Class Centroids (SRCC), a novel reweighting technique.
• Assigns sample weights based on similarities to online-learned class centroids in the feature space.
• Utilizes mixed inputs (interpolated images and labels) regularized by class centroid confidence for boundary refinement.
• Updates class centroids iteratively to learn more discriminative feature representations.

Main Results:

• SRCC demonstrates robustness to label noise by leveraging statistical class centers.
• The method effectively regularizes decision boundaries using confidence-evaluated mixed inputs.
• SRCC outperforms state-of-the-art methods on both synthetic and real-world image recognition tasks with noisy data.

Conclusions:

• SRCC offers a robust and effective solution for training deep neural networks with noisy labels.
• The proposed method enhances feature representation learning and improves model performance in the presence of label corruption.
• SRCC represents a significant advancement in learning with noisy data for image recognition.