CASIA OpenIR  > 毕业生  > 博士学位论文
时序数据的慢特征分析及其若干应用
Alternative TitleSlow Feature Analysis on Time-series Data and its Several Applications
马奎俊
Subtype工学博士
Thesis Advisor王珏
2011-01-04
Degree Grantor中国科学院研究生院
Place of Conferral中国科学院自动化研究所
Degree Discipline模式识别与智能系统
Keyword慢特征分析 不变量学习 强对流追踪 盲源信号分离 随机傅立叶特征 Slow Feature Analysis Invariance Learning Strong Convection Tracking Blind Source Separation Random Fourier Feature
Abstract时序数据是在应用中经常遇到的一类数据类型,研究对时序数据的描述与理解具有重要的实际意义。对时序数据的特征提取是研究的基础问题之一,它关系到其后续的更高级任务的成败。而在各类特征提取方法中,对时序数据的不变量学习、即提取具有不变量性质的特征,又是最重要的研究方向之一。对于时序数据来说,具有不变量性质的特征,并不是指完全不变的常值特征(因为这样的特征在时间上没有意义),而是指变化缓慢的特征。 慢特征分析就是在时序数据的非监督学习中以不变量学习为目标的一种重要算法,它的目标是对快速变化的时间序列提取变化缓慢的特征,其算法主体是求解速度协方差矩阵的特征值问题。它在很多领域都取得了成功的应用。近年来,随着应用领域的不断扩大,对算法的需求不断更新,迫切需要我们解决慢特征分析的非线性特征扩展问题。 本文对这个问题进行了初步研究,提出了对算法进行改进的可行方案,并在实验上验证了其结果的有效性。同时探讨了慢特征分析的物理解释,及其在若干问题上的应用。本文的主要工作和贡献有: 1 对慢特征分析算法从核特征扩展方面进行了改进,使算法性能得到一定程度的提升。其物理解释增强了对算法的直观理解。 2 将改进了的算法应用于对生物视觉系统的模拟,并对非监督学习下的核参数选择进行了初步的探讨,在实验上取得了更好的结果。 3 将慢特征分析应用于强对流追踪,引入了提取“动态特征”的方法,使追踪的效果得以改善。 4 将随机傅立叶特征与慢特征分析应用于盲源信号分离,设计了可以使用多核信息的信号分离算法,改善了分离结果。 总的来说,本文在对时序数据的不变量学习||即慢特征分析的理论及应用方面作了有益的探索。
Other AbstractTime-series data is a kind of data that often occurs in applications. It is very meaningful in practice to research the description and understanding of time-series data. One of the basic problems is feature extraction, which relates closely to the success of further high-level tasks. Among the di®erent kinds of feature extraction methods, invariance learning for time-series data, i.e., extraction of features which is invariant, is one of the most important research directions. For time-series data, invariant features does not mean the absolutely constant feature (because such feature doesn't have any meaning as time goes), but means feature that varies slowly. Slow feature analysis is a kind of algorithm that aims at invariance learning in unsupervised learning for time-series data. Its goal is to extract slow features for time series that varies quickly. Its main issue is to solve the eigenvalue problem of velocity covariance matrix. It has made successful applications in many fields. Recently, as the application ¯eld expands gradually and demands for the algorithm update, we urgently need to solve the problem of nonlinear features expansion for slow feature analysis. In this thesis, we studied this problem, proposed feasible solutions to improve the algorithm, and testified its validity by experiments. The main work and contribution include: 1 Improved the slow feature analysis algorithm by kernel feature expansion. The physical interpretation enhanced intuitive understanding of the algorithm. 2 Applied the improved algorithms to simulate animals' visual systems, made pilot study on selection of kernel parameters for unsupervised learning, and gained better experimental results. 3 Applied slow feature analysis to strong convection tracking, introduced methods to extract "dynamic features", and improved the tracking results. 4 Applied random Fourier features and slow feature analysis to blind source separation, designed an algorithm that could utilize multiple kernels, and gained obvious improvement for separation. In a word, in this thesis, we have made a lot of fruitful attempts and significant progresses on invariance learning or time-series data, i.e., theories and applications of slow feature analysis.
shelfnumXWLW1443
Other Identifier200618014628054
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/6316
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
马奎俊. 时序数据的慢特征分析及其若干应用[D]. 中国科学院自动化研究所. 中国科学院研究生院,2011.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20061801462805(4863KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[马奎俊]'s Articles
Baidu academic
Similar articles in Baidu academic
[马奎俊]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[马奎俊]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.