Grant-in-Aid for Scientific Research (B)
|Allocation Type||Single-year Grants|
|Research Institution||Kobe University|
ABE Shigeo Kobe University, Graduate of School of Science & Technology, Professor, 大学院・自然科学研究科, 教授 (50294195)
YOSHIMURA Motohide Kobe University, Graduate of School of Science & Technology, Assistant Professor, 大学院・自然科学研究科, 助手 (60335461)
KOTANI Manabu Kobe University, Faculty of Engineering, Associate Professor, 工学部, 助教授 (30215272)
OZAWA Seiichi Kobe University, Graduate of School of Science & Technology, Associate Professor, 大学院・自然科学研究科, 助教授 (70214129)
|Project Period (FY)
2002 – 2003
Completed(Fiscal Year 2003)
|Budget Amount *help
¥12,900,000 (Direct Cost : ¥12,900,000)
Fiscal Year 2003 : ¥6,400,000 (Direct Cost : ¥6,400,000)
Fiscal Year 2002 : ¥6,500,000 (Direct Cost : ¥6,500,000)
|Keywords||Support Vector Machine / Pattern Recognition / Diagnosis / Image Processing / Multicalss Problem / Training Method / Fuzzy Inference / Kernel Function|
We have developed multiclass support vector machines and applied them to diagnosis problems and image processing. The major results of the project are as follows :
1. Development of multiclass support vector machines
・We have developed fuzzy support vector machines that resolve unclassifiable regions in multiclass problems.
・We have developed optimal ordering of decision-tree and pairwise support vector machines to improve the generalization ability.
2. Development of fast training methods
・We have developed steepest ascent methods for pattern classification and function approximation, in which more than two data are processed at a time.
3. Evaluation for medium to large sized data sets
・We confirmed that our methods improve the generalization ability and speed up training for large sized data sets.
4. Application to diagnosis
・We have examined the feature extraction based on independent component analysis (ICA) to enhance the discrimination performance of support vector machines. We confirmed that ICA could extract the effective features from the gas leakage sound in pipes, digit patterns, and the various benchmark datasets.
・We have developed the evolutionary feature extraction using margin maximization method.
5. Application to diagnosis
・We have developed the multi-resolution feature extraction method by using 2-dimensional wavelet decomposition for the inputs of support vector machines.
・We have developed the feature extraction method as a preparation for classifying the state of protein crystals by using support vector machines.