Dynamic Loss for Robust Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Real-world datasets frequently exhibit label noise and class imbalance, hindering robust model training.
• Existing methods often tackle these issues separately, leading to performance degradation when both are present.

Purpose of the Study:

• To develop a unified approach for robust learning from datasets with both label noise and class imbalance.
• To introduce a novel meta-learning-based dynamic loss function for improved classifier performance on biased data.

Main Methods:

• A meta-learning-based dynamic loss comprising a label corrector and a margin generator was proposed.
• A hierarchical sampling strategy was employed to enrich training data with diverse and challenging samples.
• The dynamic loss components were jointly optimized via meta-learning to adapt to clean and balanced test data.

Main Results:

• The proposed method demonstrated state-of-the-art accuracy on multiple real-world and synthetic datasets.
• Experiments were conducted on datasets including CIFAR-10/100, Animal-10N, ImageNet-LT, and Webvision.
• The approach effectively handles various data biases, including long-tailed noisy data.

Conclusions:

• The novel dynamic loss effectively addresses the combined challenges of label noise and class imbalance.
• The meta-learning framework enables robust adaptation to clean and balanced data distributions.
• This work provides a significant advancement in robust learning techniques for real-world applications.