A Systematic Review on Long-Tailed Learning

Chongsheng Zhang,George Almpanidis,Gaojuan Fan,Binquan Deng,Yanbo Zhang,Ji Liu,Aouaidjia Kamel,Paolo Soda,João Gama

2024-08-01

Abstract:Long-tailed data is a special type of multi-class imbalanced data with a very large amount of minority/tail classes that have a very significant combined influence. Long-tailed learning aims to build high-performance models on datasets with long-tailed distributions, which can identify all the classes with high accuracy, in particular the minority/tail classes. It is a cutting-edge research direction that has attracted a remarkable amount of research effort in the past few years. In this paper, we present a comprehensive survey of latest advances in long-tailed visual learning. We first propose a new taxonomy for long-tailed learning, which consists of eight different dimensions, including data balancing, neural architecture, feature enrichment, logits adjustment, loss function, bells and whistles, network optimization, and post hoc processing techniques. Based on our proposed taxonomy, we present a systematic review of long-tailed learning methods, discussing their commonalities and alignable differences. We also analyze the differences between imbalance learning and long-tailed learning approaches. Finally, we discuss prospects and future directions in this field.

Machine Learning,Artificial Intelligence,Computer Vision and Pattern Recognition,Multimedia

What problem does this paper attempt to address?

The paper aims to address machine learning problems in long-tailed distribution data. Specifically:

1. **Main Objective**: To build models that perform well on long-tailed distribution datasets, particularly improving the recognition accuracy of tail (minority) samples while maintaining similar or the same accuracy for head (majority) samples.

2. **Research Scope**: Covers multiple visual learning tasks, including:
- **Object Recognition and Classification**: Improving the recognition accuracy of rare species or defective industrial products.
- **Object Detection**: Automatically locating rare objects or events in images or videos, such as equipment cracks or engineering vehicles near natural parks.
- **Image Segmentation**: Identifying and segmenting rare objects or events in images, such as pathological areas in medical imaging.

3. **Method Overview**: Proposes a new classification system that divides existing long-tailed learning methods into eight categories: data balancing, neural network design, feature enhancement, logits adjustment, loss functions, auxiliary techniques, network optimization, and post-processing techniques. Through this classification system, the latest long-tailed learning methods are systematically reviewed, and their commonalities and differences are discussed.

4. **Comparative Analysis**: Compares the differences and connections between imbalanced learning and long-tailed learning.

5. **Future Directions**: Explores the challenges and future research opportunities in this field.