吉林大学学报(理学版) ›› 2024, Vol. 62 ›› Issue (6): 1439-1446.

• • 上一篇    下一篇

基于转导推理的小样本学习方法改进

付海涛, 金晨磊, 杨亚杰, 冯宇轩   

  1. 吉林农业大学 信息技术学院, 长春 130118
  • 收稿日期:2023-12-29 出版日期:2024-11-26 发布日期:2024-11-26
  • 通讯作者: 冯宇轩 E-mail:fengyuxuan.cn@163.com

Transductive Inference Based Improvement for Few-Shot Learning

FU Haitao, JIN Chenlei, YANG Yajie, FENG Yuxuan   

  1. College of Information Technology, Jilin Agricultural University, Changchun 130118, China
  • Received:2023-12-29 Online:2024-11-26 Published:2024-11-26

摘要: 针对目前小样本图像分类推断置信度有待提高的问题, 提出一个新的结合元置信转导推理、 数据混淆方法和按特征线性调制方法的模型. 首先, 利用转导推理在训练过程中能学习到推断数据的性质, 可以有针对性地学习; 其次, 在网络结构中结合数据混淆方法, 加强对关键特征的提取, 提升模型的特征发现能力; 最后, 在转导推理框架中加入按特征线性调制变换以改进模型的小样本查询能力. 在标准数据集Mini-ImageNet和Tiered-ImageNet上进行实验的结果表明, 该模型在这两个数据集上执行5-way 1-shot任务时准确率分别提升了3.21,3.36个百分点, 在5-way 5-shot任务上准确率分别提升了2.89,1.89个百分点. 实验结果验证了该方法的有效性.

关键词: 小样本学习, 转导推理, 数据扰动, 按特征线性调制变换

Abstract: Aiming at the problem of the need to improve  confidence level in few-shot image classification inference at present, we proposed 
a new model that combined meta-confidence transductive inference, data obfuscation method, and feature-wise linear modulation method. Firstly, by using transductive inference, the model could learn properties of inference data during training process, and achieve targeted learning. Secondly, combining  data obfuscation methods  in the network architecture to enhance the extraction of key features, and  improve the feature discovery ability of the  model.  Finally, feature-wise linear modulation transformation was added to  the transductive inference framework to improve the model’s few-shot query capabilities. The results of experiments conducted on  standard datasets Mini-ImageNet and Tiered-ImageNet show  that the model improves  accuracy by  3.21 and 3.36 percentage points respectively when performing  5-way 1-shot tasks on these two datasets, and by 2.89 and 1.89 percentage points  respectively on 5-way 5-shot tasks. The experimental results validate the effectiveness of the proposed method.

Key words: few-shot learning, transductive inference, data perturbation, feature-wise linear modulation transformation

中图分类号: 

  • TP39