1999 Fiscal Year Final Research Report Summary
Pattern Recognition based on the Convex Quadratic Mathematical Programming
| Project/Area Number |
09680360
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Intelligent informatics
|
|---|
| Research Institution | Japan Advanced Institute of Science and Technology, Hokuriku |
|---|
Principal Investigator |
VLACH Milan JAIST, School of Information Science, Professor, 情報科学研究科, 教授 (40291372)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
HIROSHI Shimodaira JAIST, School of Information Science, Associate Professor, 情報科学研究科, 助教授 (30206239)
|
|---|
| Project Period (FY) |
1997 – 1999
|
| Keywords | Support Vector Machines / SVM / DTAK |
|---|
| Research Abstract |
The aim of this research is to exploit the recent sophisticated pattern classification method defined as an optimization problem whose mathematical formulation is given as either linear or convex quadratic programming. One of the benefit of such mathematical programming comes from the fact that global solution can be found whereas most of the other optimization problems suffer from local minima problem. The outcomes of this three-year research are summarized bellow.
1. Extension of GOP and SVM for multi-class problems : Since Mangasarian's GOP and Vapnik's SVM show superior classification performance for two-class problems, extension of these methods for multi-class problems is promising. The proposed extension showed better classification performance on UCI ML datasets than the classical Bayes classifier with normal density.
2. New miss-classification measure for SVM : Although SVM is thought as a realization of the Vapnik's SRM principle, the miss-classification measure used in the SVM
… More
formulation is not a good estimation of the error rate in the SRM principle. The proposed new measure gives the number of miss-classified points on the training dataset without losing the convexity of its optimization problem that assures the existence of global solutions. Experimental results showed the proposed method gives slightly better performance on the NIST handwriting digits database.
3. SVM for classification of time-series data : A new variant of SVM was developed, which can classify time series pattern such as acoustic speech data. The proposed classifier named Chain Support Vector Classifier (CSVC) consists of several SV classifiers connected in cascade. The CSVC employs an iterative learning algorithm based on Viterbi-like segmentation and the one used for the conventional SVM, though convergence has not been proven yet expect for a special case. Preliminary experiments of on-line hand-writing digits recognition showed comparable classification performance of CSVC with the single-Gaussian hidden Markov models (HMM). Besides the CSVC, another model called DTAK-SVM (dynamic time-alignment kernel SVM) was also developed. Less
|
