Sparse representation based on manifold learning
首发时间:2013-12-24
Abstract:As a technology derived from the Human Visual System, sparse coding has attracted a lot of attentions in recent years. It aims to learn sparse coordinates in terms of the basis set, which is given directly or learned from the original data set. Because of the sparsity, the learned sparse representation can be used in further data processing( such as clustering and classifying) efficiently. But the canonical sparse coding methods are all ignored the intrinsic structure of the data. From the perspective of manifold learning, this paper propose a novel sparse coding method, called Sparse Coding based on Manifold learning (MSC). Inspired by LPP, MSC finds a basis set which can be used to represent the intrinsic manifold space of the data set, and then sparse representations will be learned in this space. The most obvious advantage of MSC compared with the algorithms which impose a manifold regularizer to the objective function directly is that MSC is nonparametric. In other words, MSC is more robust. A set of evaluations on real world applications demonstrate the effectiveness of this novel algorithm.
keywords: Pattern recognition, Cluster Analysis, Sparse coding, Manifold learning.
点击查看论文中文信息
基于流形学习的稀疏表达算法
摘要:作为一种启发于人类视觉系统的技术,最近稀疏编码受到越来越多的关注。稀疏编码的主要目的是基于基集合学习新的稀疏坐标,而基通常由原始数据集直接给出或者通过原始数据集学习得到。稀疏性使得学习得到的新编码可以更加高效地应用于后续的数据处理(如聚类和分类)。但是标准的稀疏编码算法忽视了数据的本征结构。从流形学习的角度出发,本文提出了一种崭新的基于流形学习的稀疏表达算法(简称为MSC)。受保局投影算法(LPP)的启发,MSC首先学习到一组可用于描述本质流形空间的基,然后基于这组基学习稀疏编码。相比于直接添加流形正则化约束的算法,MSC的最大优点是它没有引入新的参数。换而言之,MSC更具鲁棒性。本文以一组基于现实应用的测评数据证明了此种新算法的有效性。
基金:
论文图表:
引用
No.4575298945961138****
同行评议
共计0人参与
勘误表
基于流形学习的稀疏表达算法
评论
全部评论0/1000