一种基于深度网络的少样本学习方法  

Few-shot Learning Method Based on Deep Network

作　　者：余游 冯林[1] 王格格 徐其凤 YU You;FENG Lin;WANG Ge-ge;XU Qi-feng(Department of Computer Science,Sichuan Normal University,Chengdu 610101,China)

机构地区：[1]四川师范大学计算机科学学院

出　　处：《小型微型计算机系统》2019年第11期2304-2308,共5页Journal of Chinese Computer Systems

基　　金：国家科技支撑计划项目(2014BAH11F01,2014BAH11F02)资助

摘　　要：少样本学习是目前机器学习研究领域的热点与难点.在源域和目标域分布差异很大的情况下,现有的主流少样本学习算法训练得到的模型,泛化能力较弱,导致识别率不高.针对这个问题,提出一种基于深度网络的少样本学习方法 DL-FSL(Deep Learning-based Few-Shot Learning,DL-FSL).首先,采用Bagging方法有放回地随机采样方式产生不同的训练集,针对不同的训练集,分别产生样本集、查询集.其次,建立多条异步线程,利用关系型网络学习算法以及Pytorch深度学习框架并行训练出多个不同的基模型;然后,采用概率投票方式对不同的基模型进行融合.实验结果表明,与现有方法相比,DL-FSL方法在源域和目标域分布差异很大的情况下能有效地提高少样本学习算法的识别率.Few-Shot Learning( FSL) is a hot and difficult issue in the field of machine learning. When the distribution of the source domain and the target domain are very different,the existing main FSL model has a weak generalization ability. To solve this problem,a FSL method based on Deep Network( Deep Learning-based Few-Shot Learning,DL-FSL) is proposed. Firstly,by using the random sampling method of the Bagging,the different training sets are generated. Then,using the different training sets to generate different sample sets and query sets. Secondly,a number of asynchronous threads are established. Many different base models are trained in parallel by using the relational network learning algorithm and the Pytorch deep learning framework. Finally,the different base models are ensembled by probability voting. The experimental results show that the DL-FSL method can effectively improve the recognition rate of the FSL algorithm Compared with the existing FSL method when the source and target domains have different distributions.

关 键 词：深度网络 Bagging模型 少样本学习 Pytorch 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]