概念驱动的小样本判别特征学习方法

PDF下载HTML

Title:

Concept-driven discriminative feature learning for few-shot learning

作者:

周凯锐^1,2, 刘鑫^1,2, 景丽萍^1,2, 于剑^1,2

1. 北京交通大学 交通数据分析与挖掘北京市重点实验室，北京 100044;
2. 北京交通大学 计算机与信息技术学院，北京 100044

Author(s):

ZHOU Kairui^1,2, LIU Xin^1,2, JING Liping^1,2, YU Jian^1,2

1. Beijing Key Laboratory of Traffic Data Analysis and Mining, Beijing Jiaotong University, Beijing 100044, China;
2. School of Computer Science and Technology, Beijing Jiaotong University, Beijing 100044, China

关键词:

小样本学习; 度量学习; 类表示学习; 判别特征学习; 数据增强; 图像分类; 神经网络; 深度学习

Keywords:

few-shot learning; metric learning; class representation; discriminative feature learning; data augmentation; image classification; neural network; deep learning

分类号:

TP18

DOI:

10.11992/tis.202203061

摘要:

小样本学习旨在让模型能够在仅有少量标记数据的新类中进行分类。基于度量学习的方法是小样本学习的一种有效方法，该类方法利用有标签的支持集样本构建类表示，再基于查询样本和类表示的相似性进行分类。因此，如何构建判别性更强的类表示是这类方法的关键所在。多数工作在构建类表示时，忽略了类概念相关信息的挖掘，这样容易引入样本中类别无关信息，从而降低类表示的判别性。为此本文提出一种概念驱动的小样本判别特征学习方法。该方法首先利用类别的语义信息来指导模型挖掘样本中类概念相关信息，进而构建更具判别性的类表示。其次，设计了随机掩码混合机制增加样本的多样性和识别难度，进一步提升类表示的质量。最后对处于决策边界附近的查询样本赋予更大的权重，引导模型关注难样本，从而更好地进行类表示学习。大量实验的结果表明本文提出的方法能够有效提升小样本分类任务的准确率，并且在多个数据集上优于当前先进的算法。

Abstract:

Few-shot learning (FSL) aims to recognize unlabeled samples from novel classes with few labeled samples. Metric-based methods, which obtain favorable results in FSL, construct class representations with labeled samples and classify the query samples based on the similarity between class representations and query samples. Therefore, constructing discriminative class representations is the key to these approaches. Most of the existing work ignores the mining of concept-relevant discriminative sample information when constructing class representations, which may bring noise information in samples to the class representations. To alleviate these problems, a concept-driven discriminative feature learning method tailored for FSL is proposed in this work. First, this method incorporates semantic category information to guide the mining of the class-sensitive information of labeled samples and thereby establishes a more discriminative class representation. Then, a random mask mixing mechanism is designed to increase data diversity and the identification difficulty of query samples to further improve class representation quality. Finally, it assigns higher weights to the samples near the decision boundary to guide the model to focus on difficult samples, which helps to learn better class representations. Extensive experiments show that the framework proposed in this work can effectively improve recognition accuracy, and it outperforms state-of-the-art methods on many benchmarks.