Dear Professor Badass, I on paper tried to run the SMO algorithm for an SVM problem with only 4 data points. I have the below doubts - 1) I need to run the iteration till convergence, i.e when my alphas all respect the constraint and stop changing considerably? 2) As the constraint y(w.x+b) >= 1 is already factored into the dual problem and alongside the constraint of sum (alphai x yi) = 0, how can we be certain that the SMO will respect the KKT condition of alpha_i x g(x) = 0? Because no where in the steps has this constraint been formally incorporated into the dual cost function. Hope you get to see this comment.
Thanks for your explanation of SMO in SVM, this is the best one! One question at 14:40 : when we are using kernel, how can we evaluate the stopping criterion? Because what we can compute is just X'X, but we only have X in the formula.
I see! Because we have W'X in the formula, and the expression of W contains X. Therefore, when we compute y_i(w' x_i +b), it actually has X'X in it. ;-)
At 12:51 where you have 3 quadratic functions: For the one on the right you have drawn the maximum at the right border of our range but the function still goes up inside (on the left of the right bound) our range so wouldn't then the maximum be further left?
You are absolutely right. The point is still valid that the max is at one of the boundaries of the interval. However it should have been on the left, not the right. Good catch!
Thank you very much! For now, it will only be the theory videos unfortunately. I am recording them as part of a machine learning course I am teaching at my department and there we have extra sessions for coding.
@@ritwikjain8095 Sure, you can access it here: github.com/kaspergl/ML22 Of course our students have an assigned teaching assistant, but hopefully you can still get something out of it.
These are really excellent resources, thanks Professor BadAss
Thanks! I coulnd't find explanaition on how to implement finding alphas on your own - your explanation was great!
Thanks man! Your work is greatly appreciated! ❤ from Illinois
Really good explanation thanks a lot!
Dear Professor Badass,
I on paper tried to run the SMO algorithm for an SVM problem with only 4 data points. I have the below doubts -
1) I need to run the iteration till convergence, i.e when my alphas all respect the constraint and stop changing considerably?
2) As the constraint y(w.x+b) >= 1 is already factored into the dual problem and alongside the constraint of sum (alphai x yi) = 0, how can we be certain that the SMO will respect the KKT condition of alpha_i x g(x) = 0? Because no where in the steps has this constraint been formally incorporated into the dual cost function. Hope you get to see this comment.
Thanks for your explanation of SMO in SVM, this is the best one! One question at 14:40 : when we are using kernel, how can we evaluate the stopping criterion? Because what we can compute is just X'X, but we only have X in the formula.
I see! Because we have W'X in the formula, and the expression of W contains X. Therefore, when we compute y_i(w' x_i +b), it actually has X'X in it. ;-)
@@LeyunChane Exactly!
At 12:51 where you have 3 quadratic functions:
For the one on the right you have drawn the maximum at the right border of our range but the function still goes up inside (on the left of the right bound) our range so wouldn't then the maximum be further left?
You are absolutely right. The point is still valid that the max is at one of the boundaries of the interval. However it should have been on the left, not the right. Good catch!
Great playlist, is coding video to be upload or only theory videos are here.
Thank you very much! For now, it will only be the theory videos unfortunately. I am recording them as part of a machine learning course I am teaching at my department and there we have extra sessions for coding.
@@kasperglarsen Thanks a lot for reply, should I get those coding pages, is it on github? if ! kindly share url. Thanks again
@@ritwikjain8095 Sure, you can access it here: github.com/kaspergl/ML22
Of course our students have an assigned teaching assistant, but hopefully you can still get something out of it.
@@kasperglarsen Thanks for sharing link.