国防科技大学学报2026，Vol.48Issue(3)：316-338,23.DOI:10.11887/j.issn.1001-2486.26020018
类增量学习研究进展
Recent advances in class-incremental learning
摘要
Abstract
CIL(class-incremental learning)aims to enable models to maintain discriminative ability on previously learned classes while incrementally acquiring new ones,a process in which catastrophic forgetting often occurs.This paper provided a comprehensive survey and analysis of CIL and its development trends.It clarified the definition of CIL and distinguished it from other incremental learning settings.Mainstream approaches were categorized and summarized from five perspectives:memory replay,parameter and optimization constraints,model prediction calibration,model architecture design,and transfer of pre-trained models.In addition,the commonly used evaluation metrics and datasets of CIL were reviewed,and its applications in typical vision tasks such as image generation,object detection,and semantic segmentation,as well as in emerging areas including video understanding and 3D vision were summarized.Finally,the future research directions of CIL were prospected.关键词
类增量学习/灾难性遗忘/记忆回放/参数与优化约束/模型预测校正/模型结构设计/预训练模型迁移Key words
class-incremental learning/catastrophic forgetting/memory replay/parameter and optimization constraints/model prediction calibration/model architecture design/transfer of pre-trained models分类
信息技术与安全科学引用本文复制引用
张文卓,徐昕,蒯杨柳,崔家宝,丁智勇,谢旭辉..类增量学习研究进展[J].国防科技大学学报,2026,48(3):316-338,23.基金项目
国家自然科学基金资助项目(62403485) （62403485）
