> All,
>I am surprised that very few people seem to use SVMs for commercial
>purposes.
There are two important technical obstacles for SVM. One is to train SVM
on a large dataset. The other is that testing speed is much slower than the
previous methods such as neural networks. Recently I proposed a
method to solve SVM's training on a very large dataset. For example,
I trained ten SVMs on a dataset with 1,388,000 handwritten digit
samples, 120 dimensional feature vectors on P4 1.7Ghz. The total time is
just only about 6 hours. Also, I have designed a high-performance SVM
package for research. For detail, visit the following website:
http://www.cenparmi.concordia.ca/~people/jdong/HeroSvm.html
For the testing problems, only Burges and Scholkopf reported some
good experimental results on USPS and MNIST databases with the reduced
set methods. The computational cost is very high, especially when the
number of support vectors is more than ten thousand.
For RBF kernel, this problem is related to a classical problem in applied
mathematics: fast multipole method and fast gaussian transformation.
Unfortunately both are effective only in R^2 and R^3. In a high dimensional space, the
computational cost is prohibited.
Therefore the development of a fast algorithm for SVM testing is necessary
to enable it to be popular in commercial applications.
Jianxiong Dong
>Does this have to do with the patents covering SVMs? Do these
>patents cover the application of SVMs to any problem domain? Are these
>patents enforced?
>Regards,
>Joe
--
//////////////////////////////////////////////////////////////
DONG JIANXIONG
Centre for Pattern Recognition and Machine Intelligence
GM606, Concordia University
1455 de Maisonneuve Blvd. West
Montreal Quebec H3G 1M8, Canada
Tel: (Office) (514)848-7953
Fax: (514)848-4522
E-mail: [log in to unmask]
homepage: http://www.cenparmi.concordia.ca/~people/jdong
//////////////////////////////////////////////////////////////
|