Hello,every one!
I try to use SVM to have a test with The MNIST database of handwritten
digits download from http://www.research.att.com/~yann/.I found that it
was very easy to learn all the train data correctly,but when use the
trained SVM to classify the test digits,the result is too bad.(only little
more than 50% is correct!) I think this result is understandable because I
don't incorporate the invariance in kernel,so the SVM can not distinguish
the case that a digit translating a little from a sample's place from that
a digit is completely distinct digit.
However, many papers say that they have gotten good results. From these
papers, I found that SVM have so good result(at least 90% is correct) even
with nothing extra effort to kernel type or parameters. I don't know
what's wrong with my method.
Does anybody meet tha same situation as me? Or will someone provide his
code to let me test with this MNIST database.
(Because my PC is not very fast,so I only trained the first 1000 samples
of "6"and "9" in the file "train-images-idx3-ubyten" and the first 1000
samples of "6"and "9" in " t10k-images-idx3-ubyte" as the test data.I
don't think no enough train data is the major reason of failure in my
test.)
|