Dynamic Auxiliary Soft Labels for decoupled learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Long-tailed distributions in datasets pose a significant challenge for deep learning models, particularly Convolutional Neural Networks (CNNs).
• CNNs exhibit poor performance in recognizing classes with limited sample data, a common issue in real-world datasets.

Purpose of the Study:

• To enhance the performance of deep learning models on long-tailed datasets.
• To improve the effectiveness of decoupled learning frameworks by introducing a novel soft label generation method.

Main Methods:

• Proposed Dynamic Auxiliary Soft Labels (DaSL) method using a dedicated auxiliary network to generate soft labels.
• Implemented feature-level distillation and multi-scale feature fusion to improve feature learning.
• Applied soft labels in both feature learning and classifier learning stages of a decoupled framework.

Main Results:

• DaSL effectively improves feature learning by reducing within-class variance.
• The method alleviates model overconfidence during the classifier learning stage.
• Extensive experiments on three benchmark datasets validated the superior performance of DaSL.

Conclusions:

• Dynamic Auxiliary Soft Labels (DaSL) offer a robust solution for addressing the challenges of long-tailed data in deep learning.
• The proposed method enhances decoupled learning, leading to improved accuracy and reliability in recognizing minority classes.