대용량 패턴 인식 문제를 위한 서포트 함수 기반 커널 학습 방법론
- 대용량 패턴 인식 문제를 위한 서포트 함수 기반 커널 학습 방법론
- Date Issued
- In the last decade, the kernel methods have contributed to significant
advances in research areas such as statistics, probability theory,
data mining, pattern recognition and articial intelligence with practical
success in their application to regression, classication, clustering,
ranking and visualization problems. By utilizing techniques and
theories from optimization, statistics, and most importantly functional
analysis in the feature space, kernel machines have provided principled
ways of solving these problems with
exibility, generality and accuracy.
Among various kernel methods, Support Vector Machines (SVMs) and
Gaussian Processes(GPs) have been extensively studied and applied to
diverse problems in machine learning with state-of-the-art performance.
However, as the size of data to be analyzed by these kernel machines
grows explosively due to the development of automated data
collecting, processing and storing techniques, the eciency has emerged
as a critical issue. In this circumstance, nonlinear training cost of the
kernel machines to solve quadratic programming for SVMs and eigendecomposition
problem for GPs inhibits their practical advantages. And
when the runtime complexity of the solution is a primary concern as in
many real-time applications, these kernel methods are less attractive
than the conventional linear models since the run time complexity of
kernel machines are generally proportional to the size of training data.
To this end, many researches in kernel methods have been recently devoted
to the development of fast algorithms for ecient training and
sparse kernel machines for online testing. Even though some of them
achieved fair success, most of them leads to another issues to be solved
such as local minimum, performance degrade and sensitivity to the setting.
In this thesis, we aim to provide a new class of methods which enhance
the eciency of conventional kernel machines without compensating
the performance and robustness. By utilizing important concepts
from nonlinear dynamics, the proposed methods are mainly based on the
multi-basin system constructed by kernel support functions. These include
reduced set construction method for sparser kernel machines, fast
labeling method for clustering, domain described machine for ranking
and support learning for pattern denoising. Through various experiments
on synthetic and real-world problems, we show the e
ectiveness of the proposed methods by comparing the performance with other related methods. We expect the proposed methods will pave the way
for the kernel machines to be applied eciently to the emerging largescale
applications such as high-denition image analysis, web documents
ranking and gene database clustering.
- Article Type
- Files in This Item:
- There are no files associated with this item.
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.