多视角局部距离学习及其应用

Distance learning has recently received great attention in pattern recognition and machine learning. However, existing algorithms are global distance learning and focus only on single-view dataset which means they cannot deal with multi-view dataset. This project attempts to propose a multi-view distance learning algorithm which focuses on view correlation and low dimensional manifold structure. we will propose a multi-view distance learning algorithm based on Hilbert-Schimidt independence criterion and statistical correlation. To deal with multi-view features with different probability distribution, we propose a view invariant projection based on maximum mean discrepancy, and develop an information theoretic metric learning framework which can combine the global and local geometric information within multi-view dataset. We incorporate the low rank constraint to develop a novel nonlinear metric learning base on the combination of multiple kernels of multi-view data. We will apply the proposed multi-view local learning model to the areas, including image classification, clustering and content based image retrieval. This project will try to achieve innovative progresses in the theory of multi-view distance learning model, the optimization algorithms and their applications.

近年来，距离学习在模式识别和机器学习领域得到了广泛关注,传统距离学习大多是全局距离学习算法，仅适用于单视角数据，难以直接应用于多视角数据。本项目拟研究多视角局部距离学习，主要从视角相关性及多视角数据的低维流形结构两方面展开研究，研究基于希尔伯特-施密特独立性及统计相关的多视角局部距离学习；针对各视角特征具有不同概率分布，研究基于最大均值差异的不变投影算法；研究基于信息论的融合全局及局部信息的马氏距离学习算法；为降低模型求解复杂度，研究基于低秩多核的非线性距离学习算法；将提出的多视角局部学习模型应用于图像分类、聚类以及基于内容的图像检索。力争在多视角局部距离学习模型建立、优化算法设计以及应用方面取得创新性研究成果。

距离学习在模式识别及机器学习领域得到了广泛关注，传统距离学习大多应用于单视角数据，难以应用于多视角数据。本项目研究多视角局部距离学习，主要从多视角数据相关性以及低维流形结构方面展开研究，提出了基于HSIC准则的多视角判别分析模型；提出了图正则化的跨模态哈希模型用于描述视角间以及视角内的数据几何结构特征；提出了基于混合矩阵分解的多模态数据聚类；针对多视角多标记高维数据，建立了核特征矩阵分解与最小二乘的等价关系，提出了基于最小二乘的快速特征提取算法。本项目的研究有望丰富多视角数据相似性学习的相关理论和算法，为实际应用提供理论和技术支持。