This makes me add him to the category of artists like Prof Gilbert Strang. I wonder how do you develop such skill of so lucidly explaining the most intricate concepts?
Obviously one would have to know the material inside and out but also at both a theoretical and applied level. I'm guessing he has honed his technique through years of experience with students and is able to dial into that sweet spot of just enough, but not too much, detail--while providing lots of context and interlinking of related concepts. He only goes theoretical, and judiciously at that, when it adds insight... instead of grandstanding to show off. He is truly a master at bringing students along for the ride on complex subjects.
Professor Abu-Mostafa is such a cheerful person. His explanation is very clear, but I still have to pause the video every once in a while just to have a laugh.
Very valuable addition to all ML text books, in which one can easily get drowned in the mathematics involved. He is superb to elicit the meaning of mathematics without going into the complexities of the same. Thanks for this course.
Just to reiterate what other commenters are saying.... simply excellent. I've found multiple sources and could not wrap my head around the kernel trick until I found these lectures. Abu-Mostafa separates the important concepts out from the mathematical details so that you can understand the important concepts at hand. It is easy to fill in the details later once you understand the important concept.
In slide 1 at 5:04 he talks about using SVM with non linear transform. Could someone there explain the difference between h and H? (Complex h but simple H)
I am not sure but i am pretty sure that the equation for b at minute 36:59 is wrong. When I solved I got almost the same thing except instead of y of m i got 1/(y of m ) n the same spot.
a question : he said an objective function of number of miss classification is NP hard , why ? and if it is so in soft margin SVM, the amount of violation need to be minimized and to perform this u need to check every sample whether they are violating or not, so its the same action he called np-hard. any one who knows where im wrong id be glad to hear it.
i have problem with dataSet if very small between [-1;0] and i have the approximation target between [0;1] but always the trainig performance is not go to thebest solution how can i solis problem
I am a rookie MSc student this is my first time learning these.. uh.. whatever these are... and everyone in the comments saying "Woah now i understand great explanation" etc etc. and I am just looking at the screen and feel dumb.
If you take the derivative of the lagrangian of the soft-margin SVM, with respect to w, why does Xi (the error) drop out? It should depend on w, doesnt it? i.e. different margins will have different errors. So it seems to me like a super complicated problem... Thanks for help ;)
Different margins do have different errors, but different margins do not have different Ws. That's why Xi does not depend on W. In other words, for a same hyperplane (fixed W), you can define different allowed errors (Xi).
I like this a lot for his great clarity. Except this: When you get to "Then call your quadratic programming code to hand over the alpha's", you may end up with a big can of worms, because no body seems to know how to call any of the damn quadratic programming software that is available. There seem to be hundreds of codes around with usually miserable documentation. May be left with role your own. 😁
"if i went to enter the Z space, you would have never heard from me again" haha so great
This makes me add him to the category of artists like Prof Gilbert Strang. I wonder how do you develop such skill of so lucidly explaining the most intricate concepts?
Obviously one would have to know the material inside and out but also at both a theoretical and applied level. I'm guessing he has honed his technique through years of experience with students and is able to dial into that sweet spot of just enough, but not too much, detail--while providing lots of context and interlinking of related concepts. He only goes theoretical, and judiciously at that, when it adds insight... instead of grandstanding to show off. He is truly a master at bringing students along for the ride on complex subjects.
This man answered a lot of lingering questions I had, even after reading multiple articles, papers, and watching experts on UA-cam.
lecture 14,15,16 are the best SVM videos on youtube
By far best the explanation on kernels in SVMs I found online.
An amazing lecturer. His talks are perfectly clear, insightful and interesting. Thanks for putting this online!
Professor Abu-Mostafa is such a cheerful person. His explanation is very clear, but I still have to pause the video every once in a while just to have a laugh.
Sir you are one of the best professors ever! And not just in machine learning!
Very valuable addition to all ML text books, in which one can easily get drowned in the mathematics involved. He is superb to elicit the meaning of mathematics without going into the complexities of the same. Thanks for this course.
Just to reiterate what other commenters are saying.... simply excellent. I've found multiple sources and could not wrap my head around the kernel trick until I found these lectures. Abu-Mostafa separates the important concepts out from the mathematical details so that you can understand the important concepts at hand. It is easy to fill in the details later once you understand the important concept.
These lectures are really great. Exceptionally clear and fun to watch. Thank you so much for this
Conclusion : What happens in the Z space stays in the Z space :P
Lol
you are just amazing ^^ many thanks professor, I hope if you can add more videos for more techniques such as PCA, ICA and deep learning
excellent lecture ! Best explanation I have ever seen.
"... terms will be dropping like flies" lol
Thank you Professor Yaser, it clearly explain how Kernal reduce computation.
I think you just gave me a bit of intuition of what that mysterious kernel trick is... thanks!
This has been extremely helpful. Thanks for posting!
18:17 blew my mind. LOL'ed at 26:30. All the elements of a great lecture! Excellent!
Absolutely well done and definitely keep it up!!! 👍👍👍👍👍
@45:30 Establish that the Z-space exists even if we don’t know what it is
Thank you for this great lecture. Everything explained veryvery clearly.
im the happiest guy in the world, i finally understood what's the kernel freaking trick.. thank you! شكراً جدا لحضرتك
This professor is brilliant!
and again at 41:42 Prof. Yaser says: "Third way to check that kernel is valid is 'who cares' for mercer's theorem :)"
Fantastic presentation. Thank you very much.
I burst into laughs when he described positive-semidefinite matrix in terms of a "sleeping vector" and "standing vector" at 45:00
Excellent course on Introduction to ML. Thank you professor :)
1:02:30
can't slack still be zero ? as 0*0 = 0 , so condition is still satisfied
No, because you're also maximizing beta. So, the only reason we would ever let beta be zero is when zeta is nonzero.
Mark H I dont understand
It's hilarious :) Prof. Yaser at 31:30 says: " If I had gone to Z space (Which is infinite here), you would have never heard from me again :D "
In slide 1 at 5:04 he talks about using SVM with non linear transform. Could someone there explain the difference between h and H? (Complex h but simple H)
can you explain at 1:03:18 (slide 19) that why 0
24:00 "so by doing this operation you have done an inner product in a infinite dimensional space. Congratulations!" - LOL :D
Awesome lecture, had my interest the entire time!
I am not sure but i am pretty sure that the equation for b at minute 36:59 is wrong. When I solved I got almost the same thing except instead of y of m i got 1/(y of m ) n the same spot.
Thank you for the precious lectures!!
a question : he said an objective function of number of miss classification is NP hard , why ? and if it is so in soft margin SVM, the amount of violation need to be minimized and to perform this u need to check every sample whether they are violating or not, so its the same action he called np-hard. any one who knows where im wrong id be glad to hear it.
very useful lecture. Thanks Prof ! Would love to hear more
excellent demonstration of kernel methods!
Excellent lecture. Congratulations!
You are excellent
Thank you so much, it's much better than the class I attended for Pattern Recognition!
You are Pedhanna (Big brother) from now on! Thank you!
i have problem with dataSet if very small between [-1;0] and i have the approximation target between [0;1] but always the trainig performance is not go to thebest solution how can i solis problem
@39:50 The whole idea of the kernel is that you don’t visit the Z-Space
Thanks a lot Prof. Yaser.
you are a really good lecturer!!! "Okay" :D
I am a rookie MSc student this is my first time learning these.. uh.. whatever these are... and everyone in the comments saying "Woah now i understand great explanation" etc etc. and I am just looking at the screen and feel dumb.
56:51 - I don't know, that still looks pretty complicated.
59:01 - Okay, that was pretty neat.
59:29 - Jaw hit the floor.
Brilliant teacher
Thanks Superb Lecture..
If you take the derivative of the lagrangian of the soft-margin SVM, with respect to w, why does Xi (the error) drop out?
It should depend on w, doesnt it? i.e. different margins will have different errors. So it seems to me like a super complicated problem... Thanks for help ;)
Different margins do have different errors, but different margins do not have different Ws. That's why Xi does not depend on W. In other words, for a same hyperplane (fixed W), you can define different allowed errors (Xi).
Great dude . Keep it up .
This is brilliant!
Great lecture! thank you
I like this a lot for his great clarity. Except this: When you get to "Then call your quadratic programming code to hand over the alpha's", you may end up with a big can of worms, because no body seems to know how to call any of the damn quadratic programming software that is available. There seem to be hundreds of codes around with usually miserable documentation. May be left with role your own. 😁
56:53 - did not see that coming - why was he proud over the equation?
59:46 - memorial service for beta!
Classic!
Great class! Thanks a lot!
Great lecture !
Just Brilliant!!!
讲的真好！
Fabulous！ Great Intuition!
Absolute genius !
31:20
execellent "OKs"
Ok!
The previous lecture is very helpful for understanding this: ua-cam.com/video/eHsErlPJWUU/v-deo.html
ok?
👍👍👍👍👍
We R in Z space.
OKAY
like russian
haha. great
This is almost the best explanation about kernel I find. But the tone he uses makes me really sleepy. :(
Im to stupid for this why am i here anyway
Who cares?
Great explanation!