3
$\begingroup$

I am studying image classification using SVMs and it is generally defined as so...

enter image description here

N = number of training examples

W = is the weights

f(x, W) = dot product

λ is explained to be set through cross-validation however no mention is made as to how Δ is set.

I understand that the SVM loss function wants the score of the correct class to be larger than the incorrect class scores by at least by Δ, but they don't explain how Δ is derived.

In most of the examples it is define to be Δ = 1.0, with no mention as to how 1.0 was calculated. Is this value determined through trial-and-error (cross-validation)? How does one determine what should be the value?

$\endgroup$
3
$\begingroup$

enter image description here

Intuitively, SVM wants score, xiwyi, of the correct class, yi, to be greater than any other classes, xiwj, by at least Δ such that the loss becomes zero (clamped with the max operation).

Find the full blog here. Hope it helps

$\endgroup$
5
$\begingroup$

You could use cross validation with grid search as shown here

$\endgroup$

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.