Algorithms 2025, 18(7), 422; https://doi.org/10.3390/a18070422 (registering DOI) - 9 Jul 2025
Abstract
►
Show Figures
In the process of few-shot class-incremental learning, the limited number of samples for newly introduced classes makes it difficult to adequately adapt model parameters, resulting in poor feature representations for these classes. To address this issue, this paper proposes a feature transformation method
[...] Read more.
In the process of few-shot class-incremental learning, the limited number of samples for newly introduced classes makes it difficult to adequately adapt model parameters, resulting in poor feature representations for these classes. To address this issue, this paper proposes a feature transformation method that mitigates feature degradation in few-shot incremental learning. The transformed features better align with the ideal feature distribution required by an optimal classifier, thereby alleviating performance decline during incremental updates. Before classification, the method learns a well-conditioned linear mapping from the available base classes. After classification, both class prototypes and query samples are projected into the transformed feature space to improve the overall feature distribution. Experimental results on three benchmark datasets demonstrate that the proposed method achieves strong performance: it reduces performance degradation to 24.85 percentage points on miniImageNet, 24.45 on CIFAR100, and 24.14 on CUB, consistently outperforming traditional methods such as iCaRL (44.13–50.71 points degradation) and recent techniques like FeTrIL and PL-FSCIL. Further analysis shows that the transformed features bring class prototypes significantly closer to the theoretically optimal equiangular configuration described by neural collapse, highlighting the effectiveness of the proposed approach.
Full article
Figure 1
