基于对比约束的可解释小样本学习

Interpretable Few-Shot Learning with Contrastive Constraint

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.