Ridge Regression (L2 Regularization)

[@jehushaphat]

The sign of a great teacher is the ability to make complicated concepts simple to the student. You, my friend, are a great teacher. Thank you!

[@yonko5787]

that inverse writting is f...ng awesome

[@hussameldinrabah5018]

So far after strugglıng for days, I think you make it almost clear for me about how regularization really can decline the effect of theta (or we can say the slope). I checked most of the videos about regularization, and tbh, none helped me to understand that regularization term and how it really affects the slope/steepness. You used Normal Equation to elaborate the idea of regularization that was magnificent to have a clear view about how you can decrease/decline the steepness of theta by varying lambda. The more lamba, the less steep theta be and vice versa. Unfortunately, most of videos/sources don't elaborate the intuition behind this term and how it really change the thetas/slopes. They all saying the same thing about penalizing/declining the steepness without showing why and how?

[@elgunisgandar7523]

Yeah. The same for me too.

[@prasama705]

I watched many videos about ridge regression. This is the perfect one that I have seen. Majority of the videos just simple talking about working with few parameters and doing a linear fit. You are going above that and discuss how to generalize the Ridge regression. This video is the best.

[@EndlessEngineering]

Thank you so much for watching! I am glad you found the video useful! Please let me know of there are other topics you would like to see detailed videos on

[@ahans123]

If you don't to any normalization, a reasonable choice for theta can easily be much larger than 1. Since with least squares we have a convex error surface, you don't need to normalize. However, I agree that in general normalizing your data doesn't hurt and in that case your suggestion of picking a value between 0 and 1 makes a lot of sense! Kudos for the nice explanation and derivation!

[@EndlessEngineering]

Thanks for watching! Glad you found the video enjoyable.

[@themadone7568]

Madone: This was brilliant. Its going straight from your video into Matlab. I'm beginning to understand the maths of the reservoir computing model echo location I'm writing !!!! I got this equation of ridge regression from Tholer's PhD : Wout= (Tm'*M)*((M'*M)+B*eye(N))^-1; and there at 8.51 it's derivation is explained. Thanks.

[@EndlessEngineering]

Thank you for watching, glad you found it useful

[@MrAbhishek75]

awesome...i was totally confused in ridge regression as i am new to Data science. Thanks a lot for your help.

[@EndlessEngineering]

Hi Abhishek, glad to help! Thank you for watching

[@nathancruz2843]

Great explanation. Followed along just fine after reading ISLR ridge section. Helped me see the approach of RR behind the code and text.

[@EndlessEngineering]

Thank you for watching, glad you found it useful

[@julianocamargo6674]

Very well explained. Your channel should have a lot more views.

[@EndlessEngineering]

Thank you for watching! I am glad you found it useful

[@mathhack8647]

Amazing explanation ,

[@adrianbrodowicz3485]

That's a kind of video what i was looking for. There is a lot of videos with obvious informations and nothing about mathematical representation and derivatives. You did it very well. What about constant - theta_{0}? A lot of sources say that theta_{0} shouldn't be regularized and then in the equation instead of identity matrix we should use modified identity matrix with first row full of zeros.

[@EndlessEngineering]

Hey Adrian. Thanks for watching. You bring up a good point here, and I think the answer to that is that it depends. Every model and every datasets may have different scaling requirements, and having the theta_0 (bias) term regularized or not depends on that. I have personally always implemented it with regularization, ans have not needed to take it out. I would be interested to see how that affects the results. Maybe I can do a test example and make a video on that!

[@step_by_step867]

Nice video, I appreciate it !

[@EndlessEngineering]

Thank you for watching! Glad you found it clear and useful

[@maxbarnes2240]

Excellently explained! Thank you

[@EndlessEngineering]

Thank you for watching! Glad you found it useful

[@rohitkamble8515]

Thanks for the wonderful explanation. Could you please make same video for lasso and elastic net.

[@bastiano2939]

Great explaination but why does Lambda has to be times the identity matrix?

[@EndlessEngineering]

Thank you for watching, glad you found it useful Lambda is a scalar and we can't add a scalar to a vector/matrix directly so we need to multiply by an identity matrix of the proper size

[@alanprophett9936]

Great video 😁

[@HaramChoi-l4z]

많은 도움이 되었습니다 from korea Thank you!

[@EndlessEngineering]

Thank you for watching! I am glad you found the video useful

[@Amin-gs9mn]

Thank you. You explain clearly

[@EndlessEngineering]

Thank you for watching!

[@CarpoMedia]

how can I apply this to a small artificial dataet? do you have any examples for that

[@pranjaldureja9346]

it would be superb if you could do the same from scratch in python i.e. formulation of matrices X and Y,optimizing the cost function(minima) and arriving at theta.

[@EndlessEngineering]

Hi Pranjal, I am working on this for a potential next video. Just like I did the python implementation for linear regression from scratch, I am planning for a video that does ridge regression in python. Stay tuned! And thanks for watching

[@mohamadabdulkarem206]

excellent work

[@EndlessEngineering]

Thank you! Glad you liked it!

[@usmanjavaid4195]

Nice Explaination

[@WestcoastJourney]

In the end I was like "wait what are all those formulas???"

[@EW-mb1ih]

at the beginning, why do you put a bar on top of x?

[@EndlessEngineering]

The bar is to show that this the vector x_bar = [1, x]^T. We add a 1 to the vector x to make the equation compact

[@Account-fi1cu]

Thank you, good stuff

[@EndlessEngineering]

Thank you for watching! Glad you found it helpful

[@fatmahm5787]

Thank you for you explain it was wonderful. I have a question how can i use the Ridge regression in matlab ? And if i have my input and output how will I use them in the code of ridge regression and what it will be the coefficient in ridge regression ? Please help me i can’t figure it out

[@EndlessEngineering]

Hi Fatma, thank you for your comment. I have a video on coding ridge regression in python (see link below), The code for that video is on GitHub as well. Fortunately I do not have any code in matlab, but the concepts should directly translate to a matlab implementation. kzbin.info/www/bejne/jZLXoquNe82WkM0

[@masterj2245]

Very helpful

[@EndlessEngineering]

Glad it helped! Thanks for watching

[@yatinarora9650]

What i didn't understand is , now lambda will be only on the diagonal,and how it'll help (X)TX - lamda(Identical matrix) why just the identical element , why not all

[@EndlessEngineering]

Hi Yatin. Lambda is away to penalize the model parameters from getting too large, so if you set lambda=1 you get an identity matrix. But typically in practice that is a large weight to penalize the model parameters, usually lambda is a positive number that is

[@alideeb4228]

thank you!

[@EndlessEngineering]

Thank you for watching Ali!

[@ganesh3189]

i love you so much sir

[@analistaremoto]

nice

[@anarbay24]

The final formula is not correct. You should not get identity matrix $I$ in the formula.

[@EndlessEngineering]

Hi Anar, thank you for the comment. The reason an identity matrix is required is for mathematical consistency. The first term in brackets (x^Tx) is a square matrix and we can't add a scalar (lambda) to a square matrix, so to have the correct mathematical notation the identity is required.

[@yatinarora9650]

@@EndlessEngineering why we should not have concidered at one matrix where all values are 1 instead of only diagonal as 1.

[@EndlessEngineering]

@@yatinarora9650 a diagonal of 1 does not follow the same formulation as penalizing the norm of the model parameters with lambda. In practice lambda is a positive number that is usually < 1. We do not want to penalize the norm of the model parameters too much, that might cause us to not fit the data well

[@chiaosun3505]

I love you