Recent advances in class-incremental learning

Class-incremental learning (CIL) has become an important research setting in continual learning, where a model is required to learn new classes while preserving recognition performance on previously learned classes with the label space expanding over time. In contrast to traditional supervised learning, which assumes a fixed label space and the availability of all training data at once, CIL is more consistent with the requirements of intelligent systems deployed in open and evolving environments. Since only data from the current stage are typically accessible during incremental training, model updates tend to overwrite previously acquired knowledge, leading to catastrophic forgetting. Moreover, CIL usually requires unified prediction over all encountered classes at test time without access to task identity, which makes the discrimination between previously learned and new classes particularly difficult. Since existing surveys have largely examined continual learning from a broader perspective and have not provided a focused and systematic review dedicated to CIL, a comprehensive survey of class-incremental learning is presented in this paper, covering methodological developments, evaluation protocols, task extensions, and future research directions.

A relatively well-defined methodological framework centered on catastrophic forgetting has been developed in class-incremental learning. This paper formally analyzes the CIL setting and examines the mechanisms underlying catastrophic forgetting from the perspectives of the stability-plasticity dilemma, parameter drift, gradient conflict, and representation-space degradation. Based on this analysis, existing approaches are grouped into five categories: memory replay, parameter and optimization constraints, model prediction calibration, model architecture design, and transfer of pre-trained models. Memory replay methods alleviate forgetting by storing raw exemplars, compressed representations, or generated pseudo-samples so as to approximately recover historical class distributions during incremental training. Parameter and optimization constraints methods restrict model updates through knowledge distillation, parameter regularization, and gradient constraints, thereby alleviating interference between newly acquired knowledge and previously learned classes. Methods based on model prediction calibration mainly address the prediction bias between old and new classes by recalibrating classifier outputs in the unified label space. Model architecture design methods include parameter selection and functional isolation under fixed architectures, as well as dynamic expansion strategies that increase model capacity and reduce representation conflicts across incremental stages. Transfer of pre-trained models exploits the generalization capability of large-scale pre-trained models and facilitates the transition of CIL from training-from-scratch paradigms toward pre-training-based paradigms through prompt learning, backbone fine-tuning, and transfer from vision-language models.

In addition to methodological advances, the evaluation protocols and application scope of CIL have also been reviewed. A relatively mature benchmark ecosystem has been established, with widely used datasets such as MNIST, CIFAR-100, Tiny-ImageNet, ImageNet, CUB-200, and CORe50, enabling evaluation across different dimensions including category scale, semantic complexity, scene variation, and fine-grained recognition. Commonly adopted metrics include average accuracy, average forgetting, forward transfer, and backward transfer, which together characterize knowledge retention and adaptation to new classes throughout the incremental process. Standardized codebases and experimental platforms have also become increasingly available, supporting reproducibility and fair comparison across methods. Furthermore, the applications of CIL have been reviewed not only in image classification, but also in image generation, object detection, semantic segmentation, and emerging tasks such as video understanding, 3D vision, and multimodal learning. These task extensions introduce additional challenges in output structure, spatial modeling, temporal dynamics, and cross-modal alignment, thereby extending CIL from a classification-oriented setting to a broader continual perception problem in complex intelligent systems.

Overall, substantial progress has been achieved in class-incremental learning with respect to methodological development and benchmark evaluation. However, when moving toward more realistic and dynamically changing environments, multiple challenges still remain in coping with catastrophic forgetting and adapting to newly emerging classes. As discussed in this paper, future research falls into three broad perspectives: data, strategy, and system. From the data perspective, more attention is paid to semi-supervised class-incremental learning, learning under noisy labels, and class-incremental learning with long-tailed distributions, so as to improve robustness under limited, imperfect, and imbalanced data. From the strategy perspective, it is necessary to further investigate the effective integration of multiple continual learning strategies, as well as class-incremental learning under online and dynamic task boundaries, in order to better match real-world scenarios where data arrive continuously and task boundaries are often ambiguous. From the system perspective, more efforts are still needed in the theoretical characterization of class-incremental learning, analysis of performance bounds, and the development of efficient and trustworthy CIL frameworks for practical deployment. Progress along these directions will provide stronger theoretical foundations and methodological support for building intelligent systems with the long-term adaptive capability.