7
$\begingroup$

I would like to generate samples from a support vector machine, with a Gaussian kernel and a fixed C and sigma? For examples (x1,x2) and corresponding class label y

Effectively I would like to generate a dataset such that I know a C and sigma that work well on it. Then knowing these values I can access various method of hyper-parameter optimisation methods?

I've seen how it can be done with Gaussian processes but they are probabilistic so I'm not really sure how or if it can be done with SVMs? Or is there a different way to construct a dataset to assess hyper-parameter optimisation routines?

Improve this question
$\endgroup$
1
  • $\begingroup$ Judging from your question, the hyper-parameters of the SVM are known in this case ("generate samples from a support vector machine). So, you simply have to generate random data (x1,x2), classify them with your SVM and obtain y. $\endgroup$
    – Steven
    Dec 17, 2012 at 11:17

1 Answer 1

Reset to default
5
$\begingroup$

As Steven says, I would generate new samples in a "sensible" way, and then I would use the SVM to classify them and see if they belong to the class you are aiming for.

Now, this sampling would not correspond to the true distribution of data

What do I mean by sensible? For example, in the case of object recognition or OCR, one applies some kind of transformation (3D and/or 2D rotations, warping, ...), because one expects them to happen in reality, and you want your system to be able to deal with those.

Hope it helps

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.