一种鲁棒性的少样本学习方法

少样本学习是目前机器学习研究领域的一个热点,它能在少量的标记样本中学习到较好的分类模型.但是,在噪声的不确定环境中,传统的少样本学习模型泛化能力弱.针对这一问题,提出一种鲁棒性的少样本学习方法RFSL(Robust Few-Shot Learning).首先,使用核密度估计(Kernel Density Estimation,KDE)和图像滤波(Image Filtering)方法在训练集中加入不同的随机噪声,形成多个不同噪声下的训练集,并分别生成支持集和查询集.其次,利用关系网络的关系模块通过训练集端到端地学习多个基分类器.最后,采用投票的方式对各基分类器的最末Sigmoid层非线性分类结果进行融合.实验结果表明,RFSL模型可促进小样本学习快速收敛,同时,与R-Net以及其他主流少样本学习方法相比,RFSL具有更高的分类准确率,更强的鲁棒性.

Robust Few-shot Learning Method

Few-Shot Learning is a hot topic in the field of machine learning.It can learn a better classification model from a small number of labeled samples.However,in the uncertain environment of noise,the generalization ability of traditional learning model with few samples is weak.To solve this problem,we propose a Robust Few-Shot Learning Method(RFSL).Firstly,we use the kernel density estimation(KDE)and image filtering to add different random noises into the training set.In this way,the training set under different noises is formed,and the support set and the query set are generated respectively.Secondly,the relational module of relational network is used to learn multiple base classifiers end to end through training set.Finally,we use the method of voting to merge the nonlinear classification results of the last Sigmoid layer of each base classifier.We show that with our RFSL method,our framework can promote fast convergence of Few-Shot Learning.And compared with R-Net and other classical Few-Shot Learning methods,our algorithm has higher classification accuracy and stronger robustness.