Hi all, I am currently writing a paper on comparison of SVMs, neural nets and KNN classifiers for digit recognition when the training set is very small. We have created our own dataset, with 250 training samples and 2047 testing samples. In my experiments, I observed that while linear SVM gives an accuracy of 81.44%, polynomial, RBF and sigmoid kernels give accuracy of about 20%. I am unable to understand why there is such a big difference in their performance. If somebody could clarify why this is happening, I'd be glad. -P.SatPrem Reddy