基于度量学习和知识迁移的鲁棒分类和排序学习研究

发布时间：2018-04-19 00:24

本文选题：度量学习 + 鲁棒学习　；参考：《浙江大学》2017年博士论文

【摘要】：分类和排序学习是基于数据模式判别的有监督学习问题。针对一般有监督学习应用的本质需求,分类和排序学习需要考虑如下几个因素。首先,数据在特征空间的分布中通常具有复杂的非线性几何结构。由于空间中的几何拓扑可通过一个距离度量函数刻画,我们需要学习一般的非线性距离度量函数以有效恢复数据分布的几何结构。其次,数据来源中普遍存在不可靠的噪声样本。模型需要区分特征中可靠的模式,鲁棒地反映数据的全局分布。第三,快速涌现的新领域中常常缺乏标定的训练数据。这要求模型能处理训练和目标数据具有相关但不同的特征和语义分布的情况,从训练数据中提取可适应目标数据的知识,实现有效的跨领域知识迁移。根据上述分析,本文从以下三个方向开展研究:挖掘数据分布的内在几何结构,建立学习模型的鲁棒性,和实现可适应的知识迁移。本质上,这三个方面分别针对了机器学习研究的三个基本原则,即模式挖掘的有效性、鲁棒性和可适应性。它们之间具有互相促进和补充的潜在联系。有效性关心模型拟合数据分布的精确性,关注捕捉数据分布丰富的局部非线性结构。鲁棒性关心模型拟合数据分布的全局性,关注捕捉数据的整体、全局的分布结构。可适应性关心模型提取领域间共享知识的能力,关注模型对未知领域的探索。在调研大量前人工作的基础上,本论文利用上述三个研究方向间的内在联系,提出了新颖的排序和分类学习模型,旨在实现对数据模式结构的有效挖掘,鲁棒学习和知识迁移的联合优化。本文工作取得了如下的研究成果:一,本文提出了基于Bregman距离度量函数的结构排序学习算法。Bregman距离函数族是一类有灵活、泛化的非线性形式的距离度量函数。算法利用该距离函数的非线性建模能力,学习适应数据的Bregman距离函数以有效挖掘数据分布的一般结构及语义校准。另一方面,算法基于结构学习思想建模,利用排序列表的序列结构信息以学习适应排序任务的模型。该算法提供了一个联合的距离函数学习和排序学习的一般框架,通过同时建模数据的非线性分布模式和输出列表的整体结构,实现了对排序模型的数据适应和任务适应的联合优化。二,本文提出了一个有效鲁棒的统一的分类学习框架,自步提升学习(Self-Paced Boost Learning)。该框架揭示和利用了基于有效模型选择的提升学习方法和基于鲁棒样本选择的自步学习方法的一致性和互补性,将分类模型形式化为联合的判别性模型选择和鲁棒性样本选择过程。模型通过同时从弱到强地学习分类器和从易到难地学习样本,能够在捕捉类间判别性模式的同时保证被学习样本的可靠性,实现了分类学习的有效性和鲁棒性的联合增强。三,本文提出了受语义相关性约束的可适应零样本分类模型,从知识迁移的角度研究零样本学习问题(从有训练数据的可见类别学习对无训练数据的目标类别的分类器)。模型引入了新颖的语义相关性正则化(Semantic Correlation Regularization,SCR)方法,通过约束分类器的输出符合类别间的语义相关关系,来挖掘训练类别和目标类别共同的特征和语义模式,增强模型对目标类别数据的适应性。模型基于联合的受SCR正则化的提升优化和自控制的样本选择作形式化,通过对分类器的判别性、鲁棒性和跨语义可适应性的联合增强,在零样本学习上实现了有效的知识迁移。
[Abstract]:Classification and ordination of learning is based on the data in the pattern recognition of supervised learning problems. According to the general supervised learning essence demand application, classification and ordination of learning need to consider several factors. Firstly, the distribution of data in the feature space usually has complicated structure. Because of nonlinear geometric geometry in topological spaces by a distance metric to describe, we need to learn the general nonlinear distance measure function to effectively restore the geometric structure of the data distribution. Secondly, noise samples unreliable exists in the source data. The model needs to distinguish between reliable patterns, robust reflect the distribution of the global data. Third, the training data calibration often lack the rapidly emerging new areas in this model can handle the training requirements. And the target data is related but different features and semantic distribution from the training data The extraction can be adapted to the target data to achieve cross domain knowledge, effective knowledge transfer. Based on the above analysis, this paper carried out the research from the following three aspects: the intrinsic geometric structure of distributed data mining, establish a robust learning model, and realize knowledge transfer can adapt. In essence, these three aspects respectively for three the basic principles of machine learning, namely the effectiveness of pattern mining, robustness and adaptability. With mutual promotion and complement the potential links between them. The effectiveness of care for accuracy of the model fitting the data distribution, attention capture local nonlinear structure of distributed data rich. The robustness of the model about global fitting of data distribution, overall attention capture data, global distribution structure. The adaptability of knowledge sharing model about extraction field, explore the attention model of the unknown. A lot of research in On the basis of previous work, this paper use the inner link of the three direction between the proposed classification and new learning model, to achieve effective data mining model structure, robust learning and joint optimization. This paper has made the research results such as: first, this paper puts forward a structure Bregman distance measure function ranking algorithm.Bregman distance function based on family is a kind of flexible, nonlinear generalization of the form of the distance function. The algorithm uses the function nonlinear modeling ability of the distance learning to adapt to the Bregman distance function data structure and semantic data mining effectively distributed calibration. On the other hand, the algorithm structure the thought of learning models based on the sequence structure information sorted list to learn to adapt to the sorting task model. The algorithm provides a combined distance function A general framework for learning and learning to rank number, through the whole structure while modeling data nonlinear distribution mode and the output list, realize the joint optimization of the scheduling model and the task to adapt to the data. Two, proposed a unified taxonomy effective robust learning framework, since the step (Self-Paced Boost Learning) to enhance the learning the framework reveals and use effective learning methods and model selection to enhance robust sample selection step self learning method based on consistency and complementarity, the classification model of the formal process for discriminative model selection and robustness of sample selection. At the same time combined model by learning from weak to strong classifier and from easy to difficult learning samples can also distinguish between class model to capture is to ensure the reliability of the learning samples, the classification of learning effectiveness and robustness of the joint Enhanced. Three, proposed by the semantic relevance constraints can be adapted to the zero sample classification model of zero sample from the perspective of learning knowledge transfer problems (from the visible category learning training data on the training data of the target category classifier). Introducing semantic correlation regularization of novel Correlation (Semantic Regularization, SCR) method, meet the semantic relationship between categories through the classifier output constraints, to excavate the training categories and target class common features and semantic pattern, enhance the adaptability model of target class data. By optimizing the model sample of SCR regularization and self control joint selection based on a formalization of the discriminant the combined classifier, enhance the robustness and adaptability in cross semantic, zero sample learning achieved effective knowledge transfer.

【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP181

【相似文献】