3
$\begingroup$

When you use a model that has been trained based on the RBF (Gaussian) kernel, do you need to store the entire training dataset to compute similarity features? If so, why doesn't this decrease the efficiency of RBF SVMs, since you have your data to "carry around" on top of the weights? Or do you pick the training data whose similarity functions correspond to the largest weights and only keep them?

I'm still trying to figure out what SVMs are so please let me know if the question doesn't make sense.

$\endgroup$
  • 1
    $\begingroup$ Yes - this is one of the downsides of using RBFs. There are some ways of mitigating this such as the surprise criterion, basically you only keep a training point if it changes your model significantly. There are many techniques to simplify your model that you can search through - which one is appropriate depends somewhat on your problem. $\endgroup$ – combo Sep 9 '17 at 21:34
  • $\begingroup$ I'll go ahead and post it as an answer then :) $\endgroup$ – combo Sep 10 '17 at 5:05
1
$\begingroup$

Yes - this is one of the downsides of using RBFs. There are some ways of mitigating this such as the surprise criterion, basically you only keep a training point if it changes your model significantly. There are many techniques to simplify your model that you can search through - which one is appropriate depends somewhat on your problem.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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