Lecture 15 | Machine Learning (Stanford)

[peerajak]

Sigma = (1/m)XX'

so it should be U, not V

isn't it?

I think the Lecture at 0:27:32 is not correct. I think Sigma = (1/m)XX'

but in lecture Sigma = X'X.

Anyone agree or disagree with me? please justify your opinion.

[ymwdalex]

Dose anybody think the solution is wrong? For the problem 4.3: give the PCA for natural image.

The solution is [U,S,V] = svd(X*X’); Here, X*X' is the covariance matrix. But U is not the eigenvector of the covariance matrix X.

If [U,S,V] = svd(A); U is the eigenvector of the covariance matrix of A.

[MaoMaoTM]

@ymwdalex I have the same doubt..does anyone know why the solution is svd(X*X') instead of svd(X)?

[ymwdalex]

@MaoMaoTM I mail the lecture mail address 1 month ago, but reply yet :(

[ceyceycey21]

I thought, "OMG, too many people know SVD, he is gonna skip it..". Then he didn't :D Thanks.

[cooldood83]

As some student pointed out the error at 29:39, the reason why it should be first k columns of V is the following: if X = UDV' then X'X = VDDV'=VD^2V' .... hence u shud pick the first k columns of the matrix V rather than the matrix U

[dck1760]

@cooldood83 correct



[thanhcbn]

the svd mentioned here is wrong. both U and V should be orthogonal square matrix

[xvgonzalo]

@thanhcbn should be orthonormal square matrix = unitary matrix

the svd mentioned here is the "reduced svd"

[kapildalwani]

PCA: LSI + SVD and ICA

[brunoandlydia]

ICA starts at 39:45 

[olivermitevski]

the trick at 27:00 is briliant. I would probably need to come back later to this.

[denizengineer81]

what is this reason ?