基于边际Fisher准则和迁移学习的小样本集分类器设计算法

发布时间：2018-03-30 19:49

本文选题：小样本集分类器　切入点：迁移学习　出处：《自动化学报》2016年09期

【摘要】：如何利用大量已有的同构标记数据(源域)设计小样本训练数据(目标域)的分类器是一个具有很强应用意义的研究问题.由于不同域的数据特征分布有差异,直接使用源域数据对目标域样本进行分类的效果并不理想.针对上述问题,本文提出了一种基于迁移学习的分类器设计算法.首先,本文利用内积度量的边际Fisher准则对源域进行特征映射,提高源域中类内紧凑性和类间区分性.其次,为了筛选合理的训练样本对,本文提出一种去除边界奇异点的算法来选择源域密集区域样本点,与目标域中的标记样本点组成训练样本对.在核化空间上,本文学习了目标域特征到源域特征的非线性转换,将目标域映射到源域.最后,利用邻近算法(k-nearest neighbor,k NN)分类器对映射后的目标域样本进行分类.本文不仅改进了边际Fisher准则方法,并且将基于自适应样本对筛选的迁移学习应用到小样本数据的分类器设计中,提高域间适应性.在通用数据集上的实验结果表明,本文提出的方法能够有效提高小样本训练域的分类器性能.
[Abstract]:How to use a large number of existing isomorphic tagged data (source domain) to design classifiers for small sample training data (target domain) is a significant research problem. The direct use of source domain data to classify target domain samples is not satisfactory. In view of the above problems, this paper proposes a classifier design algorithm based on migration learning. In this paper, the marginal Fisher criterion of inner product metric is used to map the source domain to improve the intra-class compactness and inter-class distinction in the source domain. Secondly, in order to select reasonable training sample pairs, In this paper, an algorithm to remove boundary singularity points is proposed to select the sample points in the dense region of the source domain and to form a training sample pair with the labeled sample points in the target domain. In the kernel space, we study the nonlinear transformation from the target domain features to the source domain features. The target domain is mapped to the source domain. Finally, we use the nearest nearest neighbor classifier to classify the target domain samples after mapping. This paper not only improves the marginal Fisher criterion method, The migration learning based on adaptive sample selection is applied to the classifier design of small sample data to improve inter-domain adaptability. The experimental results on the general data set show that, The proposed method can effectively improve the performance of small sample training domain classifier.
【作者单位】：浙江大学信息与电子工程学院;浙江大学CAD&CG国家重点实验室;
【基金】：国家自然科学基金(61471321) 教育部中国移动科研基金(MCM20150503) 国家自然科学基金(61202400) 浙江省自然科学基金(LQ12F02014)资助~~
【分类号】：TP181

【相似文献】