文献类型：期刊文献

英文题名：Semantic alignment with self-supervision for class incremental learning

作者：Fu, Zhiling[1,2];Wang, Zhe[1,2];Xu, Xinlei[1,2];Yang, Mengping[1,2];Chi, Ziqiu[1,2];Ding, Weichao[2]

机构：[1]East China Univ Sci & Technol, Minist Educ, Key Lab Smart Mfg Energy Chem Proc, Shanghai 200237, Peoples R China;[2]East China Univ Sci & Technol, Dept Comp Sci & Engn, Shanghai 200237, Peoples R China

年份：2023

卷号：282

外文期刊名：KNOWLEDGE-BASED SYSTEMS

收录：;EI(收录号：20234414993016);WOS:【SCI-EXPANDED(收录号:WOS:001109545700001)】；

基金：This work is supported by Shanghai Science and Technology Program "Federated based cross-domain and cross-task incremental learning", China under Grant No. 21511100800, Natural Science Foundation of China under Grant No. 62076094, Chinese Defense Program of Science and Technology under Grant No. 2021-JCJQ-JJ-0041, China Aerospace Science and Technology Corporation Industry-University-Research Cooperation Foundation of the Eighth Research Institute under Grant No. SAST2021-007.

语种：英文

外文关键词：Class incremental learning; Semantic alignment; Self-supervision; Self-distillation

摘要：Existing class incremental learning methods typically employ knowledge distillation to minimize discrepancies in model outputs. However, these methods are restricted by the mismatch between quondam knowledge and new data. To alleviate these issues, we introduce semantic alignment decouples the classification and distillation in different semantic spaces. The unmatched new data is regarded as out-of-distribution data on the old class distribution, and the corresponding pseudo-labels are attached to the new data using the original network. Intuitively, the pseudo-labels could be consistently preserved in the old semantic space. Moreover, we develop auxiliary self-supervised classifiers to learn more generalized representation, enabling a better stability plastic trade-off. Furthermore, self-distillation is employed to refine self-supervised knowledge from auxiliary classifiers. Extensive experiments demonstrate that our method achieves the best performance on CIFAR100, ImageNet100, ImageNet, CUB200, and Stanford-Dogs120 datasets. Notably, our method outperforms existing methods by a substantial margin when only one old exemplar is stored per class, i.e., 11.34% and 21.46% improvement on CIFAR100 of 5 phases and 10 phases, respectively.