Year: 2025
Type: article
Abstract: Most few-shot learning algorithms use complex models for better performance but at the cost of interpretability and computational effort. In recent years, some research has attempted to extrapolate the distributions of novel classes from base classes, which has attracted much attention. Methods based on this do not require complex networks in meta-training, have good interpretability, and have high classification performance. However, the existing methods only rely on the distance between class centers to evaluate the correlation between base class distributions and novel class distributions, ignoring the effect of the distribution range of base classes (i.e., the location of base class edges). Moreover, the data Gaussianization assumption relied on by existing methods does not hold in practice. We present a hypothesis about class edges and verify this hypothesis. On this theoretical basis, we propose a Center–Edge-Based Gaussianized Distribution Calibration (CGDC) method, which employs likelihood regularization and Tukey’s transformation to promote every dimension in the feature representation consistent with Gaussian distribution. More importantly, the CGDC method transfers the base class statistics to the novel classes by simultaneously considering the distribution of class centers and class edges in the feature space. Extensive experiments on miniImagenet, CUB, and CIFAR-FS datasets show that the proposed CGDC method outperforms the existing SOTA methods by an average of 1.5–2.0% in classification accuracy. In addition, the CGDC method also inherits good interpretability and low computational cost.
Source: Electronics
Authors: Huan Wan, Hantao Wang, Zhitao Ning, Xuhui Huang, Hongtian Wang, Xin Wei
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
Awards: G5460340699, G6544629769
OpenAlex ID: W4416334896