Learning From Noisy Labels With Deep Neural Networks: A Survey

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep learning models require large datasets, but real-world data often contains noisy labels.
• Noisy labels significantly degrade the generalization performance of deep neural networks.
• Robust training methods are essential for effective deep learning in practical applications.

Purpose of the Study:

• To provide a comprehensive survey of learning with noisy labels.
• To categorize and compare state-of-the-art robust training methods.
• To analyze noise rate estimation and evaluation methodologies.

Main Methods:

• Categorization of 62 robust training methods into five groups based on methodology.
• Systematic comparison of methods using six evaluation properties.
• In-depth analysis of noise rate estimation techniques.

Main Results:

• A structured overview of existing robust training techniques.
• Comparative analysis highlighting the strengths and weaknesses of different approaches.
• Summary of evaluation metrics and datasets for noisy label research.

Conclusions:

• Learning with noisy labels is a crucial area in deep learning.
• A systematic review and categorization of methods aid in understanding the field.
• Identified research gaps and future directions for robust training.