面向多源图像数据的多任务学习方法的研究

发布时间：2018-03-21 13:40

本文选题：图像分类　切入点：跨源域学习　出处：《北京交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：传统的机器学习通常假设训练数据与测试数据分布相同,现实世界这种假设往往是不成立的。传统的机器学习算法,不仅对数据的质量要求非常高,需要对数据进行去噪、降维等预处理工作,而且需要人为设计模型提取数据的特征,当样本规模非常大时,会耗费大量的人力。在大数据时代,面临着不同来源的多样性的数据集,如何将多个源域学习到的知识应用于目标域是当前跨领域学习研究的问题之一。通常不同源域的数据具有类间及类内的相关性,并且这些源域数据的分布不同,融合各源域之间的类内及类间的相关性以及其分布差异性建立适用于目标域的分类模型具有重要的意义。本文提出了基于多任务深度学习的共识正则化框架,适用于从多个源域到目标域的迁移学习。在这个框架中,每个源域采用多任务深度学习网络训练模型,将每个源域视为一个任务,并行学习多个任务不仅能提高算法的效率,而且多个任务是相关的,任务间相互约束,使得算法具有更好的分类性能。跨源域间的任务,通过引入一致性正则化项优化分类模型,使得各源域训练得到的模型不仅要考虑这个源域内的数据特性,也要考虑其它源域训练得到的模型对这个源域数据的预测,使得其它源域得到的模型与这个源域训练模型的预测保持一致。此外,本文提供了基于多任务深度学习的共识正则化框架的理论分析,并将此算法应用于图像分类领域,通过实验验证了这种学习方法的有效性。
[Abstract]:The traditional machine learning algorithm usually assumes that the training data are distributed the same as the test data, but this assumption is often not true in the real world. The traditional machine learning algorithm not only requires very high quality of the data, but also needs to de-noise the data. Dimensionality reduction and other preprocessing work, and the need to artificially design models to extract data features, when the sample size is very large, will cost a lot of manpower. In big data's time, faced with a diversity of different sources of data sets, How to apply the knowledge learned from multiple source domains to the target domain is one of the current problems in cross-domain learning. The data of different homologous domains usually have inter-class and intra-class correlations, and the distribution of these data is different. It is important to establish a classification model suitable for the target domain by combining the intra-class and inter-class correlations among the source domains and their distribution differences. In this paper, a consensus regularization framework based on multi-task in-depth learning is proposed. In this framework, each source domain uses a multi-task in-depth learning network training model to treat each source domain as a task. Learning multiple tasks in parallel can not only improve the efficiency of the algorithm, but also improve the efficiency of the algorithm. Moreover, many tasks are related, and the tasks are constrained each other, which makes the algorithm have better classification performance. By introducing the consistent regularization term to optimize the classification model, the cross-source tasks are optimized. The models obtained by each source domain training should not only consider the data characteristics in this source domain, but also consider the prediction of the source domain data obtained by other source domain training models. In addition, this paper provides a theoretical analysis of consensus regularization framework based on multi-task depth learning, and applies this algorithm to image classification. The effectiveness of this learning method is verified by experiments.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP181

【相似文献】