Jointly Optimized Classifiers for Few-Shot Class-Incremental Learning

• Published in: IEEE transactions on emerging topics in computational intelligence Vol. 8; no. 5; pp. 3316 - 3326
• Main Authors: Fu, Sichao, Peng, Qinmu, Wang, Xiaorui, He, Yang, Qiu, Wenhao, Zou, Bin, Xu, Duanquan, Jing, Xiao-Yuan, You, Xinge
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.10.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Datasets; Effectiveness; Few-shot class-incremental learning; global structure relationships; graph representation learning; Learning; local structure relationships; Message passing; multi-task training; Multitasking; Parameters; Power capacitors; Representation learning; Robustness; Task analysis; Topology; Training
• ISSN: 2471-285X; 2471-285X
• DOI: 10.1109/TETCI.2024.3375509

Few-shot class-incremental learning (FSCIL) has recently aroused widespread research interest, which aims to continually learn new class knowledge from a few labeled samples without ignoring the previous concept. One typical method is graph-based FSCIL (GFSCIL), which tends to design more complex message-passing schemes to make the classifiers' decision boundary clearer. However, it would result in poor extrapolating ability because no effort was paid to consider the effectiveness of the trained feature backbone and the learned topology structure. In this paper, we propose a simple and effective GFSCIL framework to solve the above-mentioned problem, termed Jointly Optimized Classifiers (JOC). Specifically, a simple multi-task training module incorporates both classification and auxiliary task loss to jointly supervise the feature backbone trained on the base classes. By doing so, our proposed JOC can effectively improve the robustness of the trained feature backbone, without the utilization of extra datasets or complex feature backbones. To avoid new class overfitting and old class knowledge forgetting issues of the trained feature backbone, the decouple learning strategy is adopted to fix the feature backbone parameters and only optimize the classifier parameters for the new classes. Finally, a spatial-channel graph attention network is designed to simultaneously preserve the global and local similar relationships between all classes for improving the generalization performance of classifiers. To demonstrate the effectiveness of the proposed method, extensive experiments were conducted on three popular datasets. Experimental results show that our proposed JOC outperforms many existing state-of-the-art FSCIL.