SVM with Kernels

SVM with Kernels

par Aidas Venckunas,
Number of replies: 2

Hello,

We know that in an SVM equation expressed with a kernel function we have to lambdas(langrange multipliers) are computed separately. But how exactly do we compute them? Do we have to do it by applying the theorem attached below?


Can it be asked in the exam to find lambdas? Thank you.

In reply to Aidas Venckunas

Re: SVM with Kernels

par Tsz Kin Brian Tsang,
From the lecture slides:
To solve the langrange multiplier, we need to take partial derivatives with the variables exist in the equation and set = 0 to find solution (for the equality constrained problem)

For the Inequality constrained problem, we need to make use of the KTT condition to solve, which is described in the last few slides in Optimization Basic pdf.

I also want to know whether this will be asked to compute in the exam too :)
Thank you
In reply to Aidas Venckunas

Re: SVM with Kernels

par Sena Kiciroglu,
Hi,

There are a couple more steps to go from the formulation you posted, to the "dual Lagrangian" formulation, on slide 48. This formulation is:


And the SVM algorithm aims to minimize this in terms of lambdas subject to some constraints. These lambdas are called the dual coefficients, they're the parameters of the SVM model.

During the exam, we would not ask you to do such an optimization, but it's good to know what the lambdas stand for!

Best,
Sena