【作者】 周小明；

【导师】 闵华清；

【作者基本信息】 华南理工大学 ， 计算机应用技术， 2015， 硕士

【摘要】 迁移学习作为机器学习的一个分支,由于在目标域中往往缺乏大量有标注的数据,而标注目标域数据的代价往往又十分昂贵,但是又有许多和目标域相似但又不相同的标注数据,因此就要利用这些与目标域相似的源域中的知识,将源域中有用的知识迁移到目标域内,从而来完成迁移学习。本文先介绍过去的一些相关方法,这些方法尽可能大地提高源域和目标域的相似度,来提高迁移学习的效果,来解决迁移学习中的域适应问题。目前最常用的方法就是最大均值差的方法。在论文中我们引入一种更加有效的分类均差方法,通过一组临时目标域分类,计算出源域和目标域中每个类的均值差,将这些均值差合并为一个均值差,然后同时对结构损失函数和合并后分类均差进行优化完成分类的均值差方法。更重要的是这是一种可以进行迭代下降的方法,通过迭代使得分类均值差不断下降,从而使得源域和目标域的差异越来越小,来得到尽可能优秀的迁移效果。最后利用目前迁移学习中常用的几个数据库和域适应中的一些方法进行对比实验来验证种迭代的分类均值差方法的有效性。更多还原

【Abstract】 Transfer learning is a branch of machine learning. Since labeled training samples are always scarce in target domain and it is expensive and time-consuming to label those samples, however, there are plenty of labeled samples which have some similarity with but no the same as samples of target domain, it is crucial to make advantage of the same knowledge between the two domains. So transferring those useful knowledge from source domain to target domain is the most significant in transfer learning. A lot of previous related work designed methods to close the similarity of the two domain to improve the performance of transfer learning, which is always considered to be the problem of domain adaptation. And the most popular method to solve the problem of domain adaptation is Maximum Mean Discrepancy.In this paper, we introduce Classified Mean Discrepancy which is more efficient than the traditional Maximum Mean Discrepancy. First, we should build a set of pseudo labels for target domain and then calculate the discrepancy of each class between source domain and target domain and add them together to get classified mean discrepancy. By this way, we could simultaneously optimize the structural risk functional and the classified mean discrepancy to finish the method of Classified Mean Discrepancy. More importantly, it is a method which could improve its performance by iterating itself. Via iteration, it could reduce the classified mean discrepancy and rise the accuracy. By comparing with some popular methods in domain adaptation, comprehensive experiments verify the performance of the methods of Classified Mean Discrepancy and Iterative Classified Mean Discrepancy on some popular datasets in transfer learning.更多还原