中国科学院机构知识库网格
Chinese Academy of Sciences Institutional Repositories Grid
首页 机构 成果 学者
  
  1. CAS IR Grid
  2. 地理科学与资源研究所
  3. 中国科学院地理科学与资源研究所
  4. 资源与环境信息系统国家重点实验室
  5. 外文论文
Rethinking domain-agnostic continual learning via frequency completeness learning

文献类型：期刊论文

作者Peng, Jian2,3,4; Zhang, Haitao3,4; Shen, Jing5; Li, Zeyi6; Ma, Jiayi1; Li, Haifeng7
刊名INFORMATION FUSION
出版日期2026-05-01
卷号129页码:103961
关键词Continual learning Catastrophic forgetting Domain generalization Frequency completeness learning
ISSN号1566-2535
DOI10.1016/j.inffus.2025.103961
产权排序4
文献子类Article
英文摘要Continual learning addresses knowledge acquisition while mitigating catastrophic forgetting in evolving task environments. Current spatial domain approaches exhibit limitations in cross-domain scenarios with unknown domain shifts. We reformulate cross-domain continual learning as an extension of single-domain generalization, introducing a novel frequency domain perspective that remains underexplored in continual learning research. Our analysis reveals the Forgetting Frequency Bias Hypothesis: model forgetting escalates with increasing frequency distribution gaps between tasks. Specifically, task-specific frequency overfitting emerges as a critical factor, where closer inter-task frequency distributions correlate with reduced forgetting. Building on this insight, we propose Frequency-Completeness Learning (FCL), a dual-path framework that disentangles high/lowfrequency components through spectral reconstruction to enhance frequency diversity. Complementing this, we develop Frequency Domain Shuffling (FDS), a semantic-preserving augmentation strategy that improves style diversity while maintaining domain-invariant features. Extensive experiments on incremental classification (CIFAR-100, ImageNet-100, ImageNet-R) and semantic segmentation demonstrate FCL's effectiveness. Our method achieves up to 10% improvement over baselines when integrated with existing continual learning techniques. The consistent performance gains across arbitrary domain scenarios underscore the importance of frequency completeness in addressing cross-domain continual learning challenges. The source code is available at https://github.com/GeoX-Lab/FCL.
URL标识查看原文
WOS研究方向Computer Science
语种英语
WOS记录号WOS:001637296900001
出版者ELSEVIER
源URL[http://ir.igsnrr.ac.cn/handle/311030/219378]  
专题资源与环境信息系统国家重点实验室_外文论文
通讯作者Li, Haifeng
作者单位1.Wuhan Univ, Elect Informat Sch, Wuhan 430072, Hubei, Peoples R China;
2.Univ Chinese Acad Sci, Hangzhou Inst Adv Study, Hangzhou 310024, Zhejiang, Peoples R China;
3.Tsinghua Univ, Dept Precis Instrument, Beijing 100084, Peoples R China;
4.State Key Lab Precis Space time Informat Sensing T, Beijing 100084, Peoples R China;
5.Chinese Acad Sci, Inst Geog Sci & Nat Resources Res, State Key Lab Resources & Environm Informat Syst, Beijing 100101, Peoples R China;
6.Natl Key Lab Electromagnet Space Secur, Chengdu 610036, Sichuan, Peoples R China;
7.Cent South Univ, Sch Geosci & Info Phys, Changsha 410083, Hunan, Peoples R China
推荐引用方式
GB/T 7714		 Peng, Jian,Zhang, Haitao,Shen, Jing,et al. Rethinking domain-agnostic continual learning via frequency completeness learning[J]. INFORMATION FUSION,2026,129:103961.
APA Peng, Jian,Zhang, Haitao,Shen, Jing,Li, Zeyi,Ma, Jiayi,&Li, Haifeng.(2026).Rethinking domain-agnostic continual learning via frequency completeness learning.INFORMATION FUSION,129,103961.
MLA Peng, Jian,et al."Rethinking domain-agnostic continual learning via frequency completeness learning".INFORMATION FUSION 129(2026):103961.

入库方式： OAI收割

来源：地理科学与资源研究所

浏览0
下载0
收藏0
其他版本

除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。