.png){kind=icon}
.png){kind=icon}
Return
WPE: Weighted prototype estimation for few-shot learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.imavis.2023.104757.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Few-shot learning (FSL) is a challenging task that aims to transfer a model trained on base classes with adequate labeled data to accommodate novel classes with only a few training examples. In this paper, we propose a simple but effective weighted prototype estimation (WPE) method to improve FSL. We assume that similar classes have similar distributions in the feature space so that the prototypes of novel classes can be estimated by transferring information from their similar base classes. Specifically, the proposed method learns the Gaussian-like feature distributions of similar base classes with sufficient samples and then transfers the learned distributions to cali-brate the prototype of the novel class, which is weighted by its similarities. With the estimated prototype, more robust samples can be generated to improve the FSL task. Comparative experiments are conducted to evaluate the effectiveness of our proposed algorithm on three benchmark FSL datasets. The results show that our proposed method can generate more robust samples and significantly improve FSL, outperforming the state-of-the-art methods on these datasets.
Keywords:
Few-shot learning
Knowledge transfer
Data augmentation
Prototype estimation
Image classification
Journal
IF:
4.2
Papers:
4.0K
Citations:
6.7K