A Study on Support Vector Machines
Date Issued
2007
Date
2007
Author(s)
Chen, Pai-Hsuen
DOI
en-US
Abstract
Support vector machines (SVMs) have been an effective technique for data classification.Not only it has a solid theoretical foundation, practical comparisons have also
shown that SVM is competitive with existing methods such as neural networks and decision trees. This thesis studies fast training of SVM and develops new SVM applications.
Decomposition methods are currently one of the major methods for training support vector machines. They vary mainly according to different working set selections. Existing implementations and analysis usually consider some specific selection rules. This thesis first gives a comprehensive study on Sequential Minimal Optimization (SMO)-type decomposition methods under a general and flexible way of choosing the working set. Main results include: 1) a simple asymptotic convergence proof, 2) a general explanation of the shrinking and caching techniques, and 3) the linear convergence of the method.
Extensions to some SVM variants are also discussed.
Thus, all results here apply to any SMO-type implementation whose selection rule meets the criteria of this work.
For practical implementations, SMO-decomposition methods mainly select working sets via the violation of the optimality condition, which also corresponds to first order
(i.e., gradient) information of the objective function.
Following the general framework of SMO-type methods discussed in this thesis, we develop a simple working set selection using second order information. Thus, theoretical properties such as linear convergence are established. It can be extended for indefinite kernel matrices. Experiments demonstrate that the proposed method is faster than existing selection methods using first order information.
We then apply SVMs to spinal canal segmentation, a medical image application. Past work usually requires substantial knowledge form human experts, but using SVM, we build an
automatic system. Predicted results are evaluated by human experts and classified as successful, editable or unsuccessful. We can further improve SVM classifiers by incremental learning, where new training data are added
and unnecessary ones are removed. The SVM classifier achieves 94.67% successful, 4.47% editable, and 0.96%unsuccessful rate without human interventions. Our system demonstrates that SVM is a useful tool for medical image segmentation.
Subjects
支撐向量法
資料分類
分段法
依序最佳化
醫學影像
脊柱
support vector machines
data classification
decompostion methods
sequential minimal optimization
medical images
spinal cord
Type
thesis
File(s)![Thumbnail Image]()
Loading...
Name
ntu-96-F90922008-1.pdf
Size
23.31 KB
Format
Adobe PDF
Checksum
(MD5):082171df1d84615894c72ee6782510f3