[2403.17025] Boosting Few-Shot Learning via Attentive Feature Regularization
paperAuthors
Credibility Rating
Good quality. Reputable source with community review or editorial standards, but less rigorous than peer-reviewed venues.
Rating inherited from publication venue: arXiv
This paper presents an attentive feature regularization method for few-shot learning, addressing a machine learning challenge relevant to AI systems' ability to generalize from limited data—a capability important for safe and reliable AI deployment in data-constrained scenarios.
Paper Details
Metadata
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.
Summary
This paper proposes Attentive Feature Regularization (AFR), a method to improve few-shot learning by enhancing feature representation during the mixing of samples from different categories. Rather than using simple linear interpolation, AFR employs attention mechanisms at both instance and channel levels to identify semantic relationships between categories and emphasize important feature channels. The approach achieves improved recognition accuracy on novel categories without retraining the feature extractor, particularly in 1-shot settings, and can be integrated into other few-shot learning methods.
Cited by 1 page
| Page | Type | Quality |
|---|---|---|
| Autonomous Coding | Capability | 63.0 |
Cached Content Preview
[2403.17025] Boosting Few-Shot Learning via Attentive Feature Regularization
Boosting Few-Shot Learning via Attentive Feature Regularization
Xingyu Zhu 1, 2 ,
Shuo Wang 1, 2 ,
Jinda Lu 1, 2 ,
Yanbin Hao 1, 2 ,
Haifeng Liu 3 ,
Xiangnan He 1, 2
Corresponding author
Boosting Few-Shot Learning via Attentive Feature Regularization
Xingyu Zhu 1, 2 ,
Shuo Wang 1, 2 ,
Jinda Lu 1, 2 ,
Yanbin Hao 1, 2 ,
Haifeng Liu 3 ,
Xiangnan He 1, 2
Corresponding author
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.
Introduction
In recent years, convolutional neural networks (CNNs) have demonstrated remarkable capabilities on various visual classification tasks, particularly provided with sufficient training data. However, collecting and labeling such datasets is a time-consuming and expensive procedure. As a remedy to address this challenge, few-shot learning (FSL) is proposed to classify a novel object with a scarcity of labeled data. (Ye et al. 2020 ; Peng et al. 2019 ; Wang et al. 2020 ) .
Figure 1: The analysis of manifold regularization methods.
The conventional solution of FSL involves using a CNN trained on the base categories to directly extract the global features of novel objects (Hariharan and Girshick 2017 ; Wang et al. 2018 ) . It aims to yield a transferable feature representation (textures and structures) to describe a novel category. Subsequently, these features are employed to train a classifier for recognizing novel objects. Manifold
regularization (Rodríguez et al. 2020
... (truncated, 58 KB total)5f9bc9ff5ae60ad2 | Stable ID: sid_EMu9gb9hEg