基于信息对齐的半监督少样本学习方法

Semi-supervised few-shot learning based on information alignment

针对深度学习中遇到数据样本不足,数据获取难度大的问题,提出一种基于信息对齐的半监督少样本学习方法。将支持集和查询集输入到特征提取网络得到特征向量,通过类原型计算查询集与支持集的每个局部区域对的距离用于信息对齐;采用注意力机制得到重新加权后的联合关系矩阵,利用关系模块将联合关系矩阵映射成类别的相似度分数;采用伪标签的半监督训练方法,辅助模型训练。理论和实验分析结果表明,与主流少样本学习方法相比,该方法具有更强的区分差异性的能力和更好的泛化能力。

Aiming at the problem of insufficient data samples encountered in deep learning and difficult data acquisition, a semi-supervised few-shot learning method based on information alignment(FSL-SIA) was proposed. The support set and the query set were inputted to the feature extraction network to obtain the feature vector, and the class prototype was used to calculate the distance between the query set and each local area pair of the support set for information alignment. The attention mechanism was used to obtain the re-weighted joint relationship matrix, and the relationship module was used to map the joint relationship matrix into category similarity scores. A pseudo-label semi-supervised training method was used to assist model training. The research results show that compared with mainstream few-sample learning methods, FSL-SIA has stronger ability to distinguish differences and better generalization ability.