Deep Long-Tailed Learning: A Survey.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning excels in visual recognition but struggles with imbalanced datasets (long-tailed distributions).
• Class imbalance leads to biased models, performing poorly on underrepresented (tail) classes in real-world applications.

Purpose of the Study:

• To provide a comprehensive survey of recent advancements in deep long-tailed learning.
• To categorize and review existing methods for addressing class imbalance in deep learning models.

Main Methods:

• Categorization of deep long-tailed learning methods into class re-balancing, information augmentation, and module improvement.
• Empirical analysis of state-of-the-art methods using a novel evaluation metric, relative accuracy.

Main Results:

• The survey systematically reviews various techniques designed to mitigate the negative effects of class imbalance.
• Relative accuracy is proposed as a metric to evaluate the effectiveness of methods in addressing class imbalance.

Conclusions:

• Deep long-tailed learning is a critical area with significant progress in handling visual recognition challenges.
• Identifies key applications and outlines promising avenues for future research in this rapidly evolving field.