Deep learning is a science that depends on data. Traditional deep learning methods unrealistically assume that the training models are on balanced datasets. In real-world large-scale datasets, a long-tailed distribution often occurs, with a few head classes having many samples dominating model training, while many tail classes have too few samples to be adequately learned. In recent years, the long-tailed learning has set off a research upsurge in academic circles. In this paper, we comb and analyze the literature published in high-level conferences or journals to provide a comprehensive survey of long-tailed learning.
Specifically, we categorize long-tailed learning algorithms in the field of image recognition into three main types according to the design process of deep learning models: Optimizing the sample space by enriching the quantity and semantic information of samples, optimizing the model by focusing on the four fundamental components of feature extractor, classifier, logits and loss function, and auxiliary task learning, which involves introducing auxiliary tasks to aid model training and jointly optimizing long-tailed learning models across multiple spaces. Additionally, a comprehensive comparative analysis of the strengths and weaknesses of each category is conducted based on the proposed classification method. We further extend the concept of narrow long-tail learning based on the number of samples to multi-scale generalized long-tailed learning. In addition, we briefly review long-tailed learning algorithms in other data forms, such as text data and voice data. Finally, we discuss the current challenges faced by long-tailed learning, such as poor interpretability and low data quality, and look forward to the future development directions, such as multimodal long-tailed learning and semi-supervised long-tailed learning.