项目名称: 大数据中的广义稀疏几何结构学习方法研究
项目编号: No.61300086
项目类型: 青年科学基金项目
立项/批准年度: 2014
项目学科: 自动化技术、计算机技术
项目作者: 刘日升
作者单位: 大连理工大学
项目金额: 27万元
中文摘要: 如何在"大数据"时代进行数据分析是信息科学领域需要面对的巨大挑战。稀疏几何结构学习可以有效刻画数据的本质属性及反映其聚类和分类信息。但是现有方法难以处理各种具有高维、海量特点的"大数据"及描述其分布规律。本项目拟对现有稀疏理论进行推广,提出两类广义稀疏几何结构学习新方法解决以上问题。针对传统低秩学习方法在高维、海量数据上的计算效率问题,本项目一方面从数据的隐含空间低秩性出发给出一种有效的低秩几何结构增量学习框架,另一方面针对涉及到具体模型的特点设计低复杂度计算方法并给出理论分析和基于GPU的并行化实现。针对经典稀疏表示模型无法描述数据动态分布规律这一不足,利用基本微分不变量、方向稀疏性范数和最优控制等数学工具将稀疏表示理论推广到PDE空间,通过学习微分算子字典上的稀疏表示函数来设计可以有效描述数据分布规律的PDE,并将其应用到复杂图像处理问题中。
中文关键词: 大数据;广义稀疏性;偏微分方程;图像分析;目标跟踪
英文摘要: How to perform data analysis in the era of "big data" is a huge challenge for information science. Sparse geometric structure learning can efficiently model the intrinstic properties and the clustering and classification informations of the data. However, it is difficult for existing methods to handle the so-called "big data" with high-dimensionality and large scale and reveal the dynamic data distribution. This project aims at extending classic sprase theories and provides two generalzied sparse geometric structure learning methods for solving above problems. Firstly, this project addresses the computational issues of conventional low-rank learning methods in both theoretical and algorithmic directions. On the one hand, utilizing the latent low-rank space property of the data, we provide an efficient framework for incremental low-rank geometric structure learning. On the other hand, we design fast numerical algorithm for related models, study the corresponding convergence theory and the GPU based implementation. Secondly, to address the weakness of classic sparse representation for modeling the dynamic distribution of the data, we utilize several mathematical tools, such as fundamental differential invariants, directional sparsity norm and optimal control, to extend sparse representation for the space of PDE.
英文关键词: Big Data;Generalized Sparsity;PDEs;Image Analysis;Object Tracking