Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Boosting Few-Shot Learning via Attentive Feature Regularization

Xingyu Zhu, Shuo Wang, Jinda Lu, Yanbin Hao, Haifeng Liu, Xiangnan He

TL;DR

Few-shot learning often relies on manifold regularization to mix samples, but naive mixing can degrade feature quality. The paper introduces Attentive Feature Regularization (AFR), which uses semantic relations to select related base categories and applies instance- and channel-level attention to emphasize complementary and discriminative features, guided by three losses: $ ext{L}_{CE}$, $ ext{L}_{SC}$, and $ ext{L}_{MSE}$. AFR operates on pre-trained features, enabling seamless integration with existing FSL methods and classifiers, and demonstrates state-of-the-art performance on Mini-ImageNet, Tiered-ImageNet, and Meta-Dataset, especially in the 1-shot setting. The results indicate AFR’s effectiveness in improving feature representativeness and discriminability without retraining the feature extractor, with future work exploring graph-based enhancements like GCN/GNN to further boost performance.

Abstract

Few-shot learning (FSL) based on manifold regularization aims to improve the recognition capacity of novel objects with limited training samples by mixing two samples from different categories with a blending factor. However, this mixing operation weakens the feature representation due to the linear interpolation and the overlooking of the importance of specific channels. To solve these issues, this paper proposes attentive feature regularization (AFR) which aims to improve the feature representativeness and discriminability. In our approach, we first calculate the relations between different categories of semantic labels to pick out the related features used for regularization. Then, we design two attention-based calculations at both the instance and channel levels. These calculations enable the regularization procedure to focus on two crucial aspects: the feature complementarity through adaptive interpolation in related categories and the emphasis on specific feature channels. Finally, we combine these regularization strategies to significantly improve the classifier performance. Empirical studies on several popular FSL benchmarks demonstrate the effectiveness of AFR, which improves the recognition accuracy of novel categories without the need to retrain any feature extractor, especially in the 1-shot setting. Furthermore, the proposed AFR can seamlessly integrate into other FSL methods to improve classification performance.

Boosting Few-Shot Learning via Attentive Feature Regularization

TL;DR

Few-shot learning often relies on manifold regularization to mix samples, but naive mixing can degrade feature quality. The paper introduces Attentive Feature Regularization (AFR), which uses semantic relations to select related base categories and applies instance- and channel-level attention to emphasize complementary and discriminative features, guided by three losses: , , and . AFR operates on pre-trained features, enabling seamless integration with existing FSL methods and classifiers, and demonstrates state-of-the-art performance on Mini-ImageNet, Tiered-ImageNet, and Meta-Dataset, especially in the 1-shot setting. The results indicate AFR’s effectiveness in improving feature representativeness and discriminability without retraining the feature extractor, with future work exploring graph-based enhancements like GCN/GNN to further boost performance.

Abstract

Few-shot learning (FSL) based on manifold regularization aims to improve the recognition capacity of novel objects with limited training samples by mixing two samples from different categories with a blending factor. However, this mixing operation weakens the feature representation due to the linear interpolation and the overlooking of the importance of specific channels. To solve these issues, this paper proposes attentive feature regularization (AFR) which aims to improve the feature representativeness and discriminability. In our approach, we first calculate the relations between different categories of semantic labels to pick out the related features used for regularization. Then, we design two attention-based calculations at both the instance and channel levels. These calculations enable the regularization procedure to focus on two crucial aspects: the feature complementarity through adaptive interpolation in related categories and the emphasis on specific feature channels. Finally, we combine these regularization strategies to significantly improve the classifier performance. Empirical studies on several popular FSL benchmarks demonstrate the effectiveness of AFR, which improves the recognition accuracy of novel categories without the need to retrain any feature extractor, especially in the 1-shot setting. Furthermore, the proposed AFR can seamlessly integrate into other FSL methods to improve classification performance.
Paper Structure (20 sections, 11 equations, 4 figures, 5 tables)

This paper contains 20 sections, 11 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Knowledge Transfer in Few-Shot Learning
  4. Regularization in Few-Shot Learning
  5. Attention in Few-Shot Learning
  6. Method
  7. Preliminaries
  8. Attentive Feature Regularization
  9. Instance Attention
  10. Channel Attention
  11. Training and Inference
  12. Experiments
  13. Experimental Settings
  14. Ablation Study
  15. The influences of semantic selection (RQ1)
  16. ...and 5 more sections

Figures (4)

  • Figure 1: The analysis of manifold regularization methods.
  • Figure 2: The overview of attentive feature regularization (AFR), where $\mathcal{L}_{\rm CE}$, $\mathcal{L}_{\rm SC}$, and $\mathcal{L}_{\rm MSE}$ are three losses.
  • Figure 3: The calculation of calibration in instance attention.
  • Figure 4: The accuracy (%) of the classifiers trained with the different numbers of selected base categories.