The paper introduces the concept of Open Long-Tailed Recognition (OLTR), which involves learning from naturally distributed data with a long-tailed and open-ended distribution, and optimizing classification accuracy over a balanced test set that includes head, tail, and open classes. OLTR addresses challenges such as imbalanced classification, few-shot learning, and open-set recognition in a single integrated algorithm. The key contributions include: 1. **Definition of OLTR**: Defining OLTR as a task that learns from long-tail and open-end distributed data and evaluates classification accuracy over a balanced test set. 2. **Dynamic Meta-Embedding**: Developing an integrated OLTR algorithm that maps images to a feature space using a dynamic meta-embedding, which combines a direct image feature and an associated memory feature. The feature norm indicates familiarity with known classes. 3. **Modulated Attention**: Introducing modulated attention to maintain discrimination between head and tail classes by encouraging different spatial contexts for different classes. 4. **Experiments**: Curating three large-scale OLTR datasets (ImageNet-LT, Places-LT, and MS1M-LT) and evaluating the proposed method on these datasets, demonstrating superior performance compared to state-of-the-art methods. The paper also provides a detailed methodology, experimental setup, and analysis of the effectiveness of the proposed approach. The code, datasets, and models are publicly available for future research.The paper introduces the concept of Open Long-Tailed Recognition (OLTR), which involves learning from naturally distributed data with a long-tailed and open-ended distribution, and optimizing classification accuracy over a balanced test set that includes head, tail, and open classes. OLTR addresses challenges such as imbalanced classification, few-shot learning, and open-set recognition in a single integrated algorithm. The key contributions include: 1. **Definition of OLTR**: Defining OLTR as a task that learns from long-tail and open-end distributed data and evaluates classification accuracy over a balanced test set. 2. **Dynamic Meta-Embedding**: Developing an integrated OLTR algorithm that maps images to a feature space using a dynamic meta-embedding, which combines a direct image feature and an associated memory feature. The feature norm indicates familiarity with known classes. 3. **Modulated Attention**: Introducing modulated attention to maintain discrimination between head and tail classes by encouraging different spatial contexts for different classes. 4. **Experiments**: Curating three large-scale OLTR datasets (ImageNet-LT, Places-LT, and MS1M-LT) and evaluating the proposed method on these datasets, demonstrating superior performance compared to state-of-the-art methods. The paper also provides a detailed methodology, experimental setup, and analysis of the effectiveness of the proposed approach. The code, datasets, and models are publicly available for future research.