A Time-Consistency Curriculum for Learning From Instance-Dependent Noisy Labels.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Machine learning algorithms often struggle with noisy labels, especially when noise is instance-dependent.
• Real-world noisy labels arise from complex, instance-specific mechanisms, posing a significant challenge to model robustness.

Purpose of the Study:

• To address the challenge of instance-dependent noisy labels in machine learning.
• To develop a novel end-to-end training approach for robust classification from noisy data.

Main Methods:

• Jointly training a model to reverse the noise-generating mechanism and a robust classifier.
• Utilizing the instance-dependent mapping between clean and noisy labels.
• Developing a time-consistency curriculum to select informative training data based on model output dynamics.

Main Results:

• The proposed method demonstrates robust classification performance on instance-dependent noisy labels.
• The time-consistency curriculum significantly enhances training by providing high-quality data.
• Outperforms state-of-the-art methods in challenging noisy label settings.

Conclusions:

• Joint training of a noise-reversing model and a robust classifier is effective for instance-dependent noisy labels.
• Data time-consistency is crucial for successful training, and a curriculum based on it improves performance.
• The approach offers a promising solution for real-world machine learning applications with inherent data noise.