Introduction To Support Vector Machines

Institute for Signal and Information Processing
Mississippi State University
Department of Electrical and Computer Engineering
phone/fax: 601-325-3149; email:
URL: /research/isip/resources

We have seen during our previous seminars how classical statistical techniques like Linear Discriminant Analysis (LDA) and Principal Component Analysis are used in solving classification problems. They however have restricted performance when the decision surfaces required are non-linear. So, when data sets are inherently separated by a non-linear decision surface, it is advantageous to use a non-linear classifier. Neural Networks and K-NN classifier are some of the widely used non-linear classifiers.

In this talk, we introduce a new classification scheme called the Support Vector Machine (SVM) which has gained prominence in the past couple of years. It has some interesting features, like control over generalizability, maximum expected error and most importantly has a discriminative rather than a representative character. We present preliminary results of using SVMs to classify 11 vowels on a standard non-linear classification data set.