Modifications of kernels to improve support vector machine classifiers

Kernel function is a key factor in support vector machine classifiers. We put forward a new conformal transformation on kernel functions to improve the performance of support vector machine classifiers, which is based on the method of Amari's idea. We have some important modifications and make the method more robust with respect to the input data distribution and have greater generalization ability with noisy data. We have also studied the performance of the modified kernels on the Gaussian RBF and polynomial kernels when a kernel is modified iteratively several times. Simulation results for the data set comparing to the two former methods show remarkable improvement in generalization errors.