CMW-Net: Learning a Class-Aware Sample Weighting Mapping for Robust Deep Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks (DNNs) are prone to overfitting due to biased training data, including corrupted labels and class imbalance.
• Existing sample re-weighting methods often require manual specification of weighting schemes, limiting their practical applicability.
• Data bias presents significant complexities and inter-class variations, necessitating adaptive solutions.

Purpose of the Study:

• To develop a meta-model capable of adaptively learning explicit sample weighting schemes directly from data.
• To address the limitations of manual, problem-specific weighting methods in deep learning.
• To enhance the robustness of DNNs against various forms of data bias.

Main Methods:

• Proposed a meta-model that treats each training class as a distinct learning task.
• Developed an explicit weighting function that takes sample loss and task/class features as input to output sample weights.
• Implemented adaptively varying weighting schemes tailored to the intrinsic bias characteristics of different sample classes.

Main Results:

• Demonstrated the method's capability in achieving effective weighting schemes for class imbalance, feature-independent and dependent label noise, and complex bias scenarios.
• Validated the task-transferability of the learned weighting scheme by deploying a model trained on CIFAR-10 to the WebVision dataset.
• Confirmed the method's general applicability to robust deep learning problems like partial-label learning, semi-supervised learning, and selective classification.

Conclusions:

• The proposed meta-model effectively learns adaptive sample weighting schemes, significantly improving DNN robustness against data bias.
• The approach offers a general and transferable solution for various data bias challenges in deep learning.
• The method provides a flexible framework for enhancing model performance in real-world, data-rich environments.