区别性知识利用的迁移分类学习
发布时间：2018-10-05 21:53
【摘要】：目前的迁移学习模型旨在利用事先准备好的源域数据为目标域学习提供辅助知识,即从源域抽象出与目标域共享的知识结构时,使用所有的源域数据。然而,由于人力资源的限制,收集真实场景下整体与目标域相关的源域数据并不现实。提出了一种泛化的经验风险最小化选择性知识利用模型,并给出了该模型的理论风险上界。所提模型能够自动筛选出与目标域相关的源域数据子集,解决了源域只有部分知识可用的问题,进而避免了在真实场景下使用整个源域数据集带来的负迁移效应。在模拟数据集和真实数据集上进行了仿真实验,结果显示所提算法较之传统迁移学习算法性能更佳。域相关的源域数据并不现实。提出了一种泛化的经验风险最小化选择性知识利用模型,并给出了该模型的理论风险上界。所提模型能够自动筛选出与目标域相关的源域数据子集,解决了源域只有部分知识可用的问题,进而避免了在真实场景下使用整个源域数据集带来的负迁移效应。在模拟数据集和真实数据集上进行了仿真实验,结果显示所提算法较之传统迁移学习算法性能更佳。
[Abstract]:The current transfer learning model aims to provide auxiliary knowledge for target domain learning using pre-prepared source domain data, that is, when abstracting the knowledge structure shared with the target domain from the source domain, all the source domain data are used. However, due to the limitation of human resources, it is not realistic to collect the source domain data related to the target domain in the real scene. In this paper, a generalized empirical risk minimization selective knowledge utilization model is proposed, and the upper bound of the theoretical risk of the model is given. The proposed model can automatically filter out the subset of source domain data related to the target domain, which solves the problem that only part of the knowledge is available in the source domain, and thus avoids the negative migration effect brought by using the whole source domain data set in the real scene. Simulation experiments on simulated data sets and real data sets show that the proposed algorithm performs better than the traditional migration learning algorithm. Domain-related source domain data is not realistic. In this paper, a generalized empirical risk minimization selective knowledge utilization model is proposed, and the upper bound of the theoretical risk of the model is given. The proposed model can automatically filter out the subset of source domain data related to the target domain, which solves the problem that only part of the knowledge is available in the source domain, and thus avoids the negative migration effect brought by using the whole source domain data set in the real scene. Simulation experiments on simulated data sets and real data sets show that the proposed algorithm performs better than the traditional migration learning algorithm.
【作者单位】： 江南大学数字媒体学院;
【基金】：国家自然科学基金No.61272210~~
【分类号】：TP181
[Abstract]:The current transfer learning model aims to provide auxiliary knowledge for target domain learning using pre-prepared source domain data, that is, when abstracting the knowledge structure shared with the target domain from the source domain, all the source domain data are used. However, due to the limitation of human resources, it is not realistic to collect the source domain data related to the target domain in the real scene. In this paper, a generalized empirical risk minimization selective knowledge utilization model is proposed, and the upper bound of the theoretical risk of the model is given. The proposed model can automatically filter out the subset of source domain data related to the target domain, which solves the problem that only part of the knowledge is available in the source domain, and thus avoids the negative migration effect brought by using the whole source domain data set in the real scene. Simulation experiments on simulated data sets and real data sets show that the proposed algorithm performs better than the traditional migration learning algorithm. Domain-related source domain data is not realistic. In this paper, a generalized empirical risk minimization selective knowledge utilization model is proposed, and the upper bound of the theoretical risk of the model is given. The proposed model can automatically filter out the subset of source domain data related to the target domain, which solves the problem that only part of the knowledge is available in the source domain, and thus avoids the negative migration effect brought by using the whole source domain data set in the real scene. Simulation experiments on simulated data sets and real data sets show that the proposed algorithm performs better than the traditional migration learning algorithm.
【作者单位】： 江南大学数字媒体学院;
【基金】：国家自然科学基金No.61272210~~
【分类号】：TP181
【参考文献】
相关期刊论文 前3条
1 杭文龙;蒋亦樟;刘解放;王士同;;迁移近邻传播聚类算法[J];软件学报;2016年11期
2 毛发贵;李碧雯;沈备军;;基于实例迁移的跨项目软件缺陷预测[J];计算机科学与探索;2016年01期
3 杨昌健;邓赵红;蒋亦樟;王士同;;基于迁移学习的癫痫EEG信号自适应识别[J];计算机科学与探索;2014年03期
【共引文献】
相关期刊论文 前6条
1 龚凯乐;成颖;孙建军;;基于参与者共现分析的博文聚类研究[J];现代图书情报技术;2016年10期
2 谢丽潇;邓赵红;史荧中;王士同;;面向癫痫EEG自适应识别的迁移径向基神经网络[J];计算机科学与探索;2016年12期
3 程e，
本文编号：2255002
本文链接：https://www.wllwen.com/guanlilunwen/renliziyuanguanlilunwen/2255002.html
