面向数据表示的深度稀疏保持学习

项目名称： 面向数据表示的深度稀疏保持学习

项目编号： No.61300154

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： 乔立山

作者单位： 聊城大学

项目金额： 23万元

中文摘要： 数据表示（或特征）通常对模式识别算法的最终性能产生决定性的影响。传统的数据表示方法（如核技巧、流形学习等）一般依赖于局部性假设，尽管其在大量实际问题中发挥着重要作用，但往往面临维数灾难等一系列挑战。近年来，以编码或保持数据稀疏结构为目标的非局部特征学习技术在很多高维数据集上获得了经验性的成功，引起国内外研究者的关注。然而，单纯的稀疏性原则尚不足以表示具有复杂结构的数据（如自然图像），为此，本项目拟拓展传统的稀疏保持学习技术，并进行如下深入研究：1）基于多任务学习框架构建同时编码局部与非局部信息的广义稀疏性保持投影算法（已初步完成），并借此探讨局部性与稀疏性的本质区别和内在联系；2）设计深度稀疏保持学习算法，提高稀疏保持策略对复杂结构数据的特征表示能力；3）尝试探索上述特征学习算法的统计基础，为稀疏性原则和深度的分层学习结构（两者被认为与人脑的工作机理联系密切）提供更多可能的理论依据。

中文关键词： 模式识别；数据表示；特征学习；稀疏表示；功能脑网络

英文摘要： Data representation (or features) generally has a decisive impact on the final performance of patter recognition algorithms. Traditional data representation, such as kernel trick and manifold learning, usually depends on locality assumption which plays an important role in many practical problems, but faces a series of challenges including the curse of dimensionality. In the recent years, the non-local feature learning techniques achieve empirical success on many high-dimensional data by encoding or preserving the sparse structure among data,and thus cause wide concern of researchers at home and abroad. However, sparsity principle alone is insufficient to represent the data with complex structure (e.g., natural images). In this project, we plan to make a deep study of sparsity preserving feature learning, including 1) developing generalized sparsity preserving projections algorithm encoding local and non-local information simultaneously based on multi-task learning framework, which is expected to be useful for discussing the differences and relations between the locality and sparsity principles; 2) designing deep sparsity preserving learning algorithm to improve the capability of sparsity preserving strategy for representing features of the data with complex structures; 3) attempting to exploring the statistical

英文关键词： Pattern Recognition；Data Representation；Feature Learning；Sparse Representation；Functional Brain Network