基于对比约束的可解释小样本学习  

Interpretable Few-Shot Learning with Contrastive Constraint

作　　者：张玲玲[1] 陈一苇 吴文俊 魏笔凡[1] 罗炫 常晓军 刘均[1] Zhang Lingling;Chen Yiwei;Wu Wenjun;Wei Bifan;Luo Xuan;Chang Xiaojun;Liu Jun(School of Computer Science and Technology,Xi'an Jiaotong University,Xi'an 710049;School of Computing Technologies,Royal Melbourne Institute of Technology University,Melbourne,Australia 3000)

机构地区：[1]西安交通大学计算机科学与技术学院,西安710049 [2]皇家墨尔本理工大学计算技术学院,澳大利亚墨尔本3000

出　　处：《计算机研究与发展》2021年第12期2573-2584,共12页Journal of Computer Research and Development

基　　金：国家重点研发计划项目(2020AAA0108800);国家自然科学基金项目(62137002,61937001,62176209,62176207,62106190,62050194);国家自然科学基金创新群体(61721002);教育部创新团队(IRT_17R86);基于MOOC中国的“一带一路”人才培养的线上线下混合教学支撑信息化平台与服务体系;中国博士后面上项目(2020M683493);中国工程科技知识中心项目;中央高校基本科研项目(xhj032021013-02)。

摘　　要：不同于基于大规模监督的深度学习方法,小样本学习旨在从极少的几个样本中学习这类样本的特性,其更符合人脑的视觉认知机制.近年来,小样本学习受到很多学者关注,他们联合元学习训练模式与度量学习理论,挖掘查询集(无标记样本)和支持集(少量标记样本)在特征空间的语义相似距离,取得不错的小样本分类性能.然而,这些方法的可解释性偏弱,不能为用户提供一种便于直观理解的小样本推理过程.为此,提出一种基于区域注意力机制的小样本分类网络INT-FSL,旨在揭示小样本分类中的2个关键问题:1)图像哪些关键位置的视觉特征在决策中发挥了重要作用;2)这些关键位置的视觉特征能体现哪些类别的特性.除此之外,尝试在每个小样本元任务中设计全局和局部2种对比学习机制,利用数据内部信息来缓解小样本场景中的监督信息匮乏问题.在3个真实图像数据集上进行了详细的实验分析,结果表明:所提方法INT-FSL不仅能有效提升当前小样本学习方法的分类性能,还具备良好的过程可解释性.Different from deep learning with large scale supervision,few-shot learning aims to learn the samples'characteristics from a few labeled examples.Apparently,few-shot learning is more in line with the visual cognitive mechanism of the human brain.In recent years,few-shot learning has attracted more researchers'attention.In order to discover the semantic similarities between the query set(unlabeled image)and support set(few labeled images)in feature embedding space,methods which combine meta-learning and metric learning have emerged and achieved great performance on few-shot image classification tasks.However,these methods lack the interpretability,which means they could not provide a reasoning explainable process like human cognitive mechanism.Therefore,we propose a novel interpretable few-shot learning method called INT-FSL based on the positional attention mechanism,which aims to reveal two key problems in few-shot classification:1)Which parts of the unlabeled image play an important role in classification task;2)Which class of features reflected by the key parts.Besides,we design the contrastive constraints on global and local levels in every few-shot meta task,for alleviating the limited supervision with the internal information of the data.We conduct extensive experiments on three image benchmark datasets.The results show that the proposed model INT-FSL not only could improve the classification performance on few-shot learning effectively,but also has good interpretability in the reasoning process.

关 键 词：小样本学习 可解释性分析 对比学习 局部描述子 图像识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]