项目名称: 回归函数梯度的随机逼近快速算法研究及应用
项目编号: No.11201420
项目类型: 青年科学基金项目
立项/批准年度: 2013
项目学科: 数理科学和化学
项目作者: 董雪梅
作者单位: 浙江工商大学
项目金额: 22万元
中文摘要: 回归函数的梯度能够同时提供高维数据特征变量选择和维数约简信息。本项目拟综合利用函数逼近论、概率论、泛函分析和最优化理论中的相关方法,设计流形假设下半监督函数梯度学习算法,研究其推广性能;研究及相依同分布数据下的函数梯度学习算法的一致性与推广误差估计;研究球面上的核函数构造与性质,对基于该类型数据的梯度学习进行误差分析。在以上研究的基础上发展和完善基于不同类型数据的回归函数梯度逼近理论,设计快速学习算法,建立数据分布规律与学习算法的推广性能之间的关系。将相关算法应用于遥感图像数据和三维人脸数据的维数约简中。
中文关键词: 学习算法;随机逼近;正则化;;
英文摘要: The gradient of regression function can provide feature variable selection and dimension reduction information for high dimensional data at the same time. This project will utilize related methods in function approximation theory, probability theory, functional analysis, and optimization theory, etc., to design semi-supervised learning algorihms for the gradient under manifold assumption and study their genaralization properties. Consider the consistency and genaralization error of learning gradient algorithms based on dependent and identical data. Study construction and properties of kernel functions based on spherical data and analysis the generalization error of the gradient learning algorithms from this kind of data. On the basis of the above studies,develop and improve the approximation theory for gradient function based on different types of data, design fast learning algorithms, establish the relationship between the data distribution and the generalization performance of the learning algorithms.Using related algorithms to the dimension reduction for remote sensing image data and 3D face data.
英文关键词: learing algorithm;random approximation;regularization;;