多核学习若干关键问题研究

项目名称： 多核学习若干关键问题研究

项目编号： No.61272198

项目类型： 面上项目

立项/批准年度： 2013

项目学科： 自动化技术、计算机技术

项目作者： 王喆

作者单位： 华东理工大学

项目金额： 80万元

中文摘要： 本项目是对多核分类学习中遗留的若干关键问题展开的深入研究与探讨。多核学习是当前机器学习的一个新的热点，其已基本解决了多核构造理论、解的稀疏性与实际应用转化等关键问题，然而其遗留如下问题仍待解决。首先，多核构造方式仍较为单一。本项目不同于传统多核构造及权重的设置方式，在充分考虑分类模型效率的前提下，设计出一种新的多核构造方法。第二，计算代价较大。本项目从降低多核模型的计算代价与提高推广性能两个角度出发，考虑实际问题中数据的规模，使得所设计的模型应用于大规模数据时能在时空复杂度上有所突破。第三，缺乏先验信息的充分融合。本项目通过融入数据的先验信息以提高多核模型的推广性能。最后，本项目将所设计的多核模型拓展至子空间学习，设计出能去除多种复杂噪声的多核子空间学习算法，并将上述研究成果应用到基于红外图像人脸识别的实用系统中。本项目力图为分类学习理论及算法提供新的设计理念，并可望取得实际的应用成果。

中文关键词： 多核学习；经验核映射；计算复杂度；先验结构信息；模式识别

英文摘要： This proposal gives a further research on the key problems left in multiple kernel learning (MKL). As a focused problem in machine learning, MKL has basically solved the problems about multiple kernel construction, sparse solutions, real-application transformation, etc. However, MKL has still some key problems to be further discussed. First of all, it is singler for the existing construction of multiple kernels. Differently from the existing ways of constructing and optimizing kernels, this proposal develops a novel way of constructing and optimizing kernels so as to bring a high classification effectiveness. Secondly, it is larger for the computational cost of the existing MKL. Thus this proposal analyzes the scale of the dealt data and presents an efficient and effective MKL that could not only reduce the computational cost but also improve the generalization ability. Thirdly, the existing MKL lacks the enough integration by the prior information which should be fully induced in classification processing. This proposal aims to introduce the prior information into MKL so as to get a high generalization. Finally, this proposal generalizes the MKL into the sub-space learning and proposes a robust sub-space learning based on multiple kernels. Doing so could remove different noises since multiple kernels with diffe

英文关键词： Multi-kernel learning；Empirical kernel projection；Computational complexity；Prior structural information；Pattern recognition