COMS 4721: Machine Learning for Data Science Lecture 19, 4/6/2017

Transcription

1 COMS 4721: Machine Learning for Data Science Lecture 19, 4/6/2017 Prof. John Paisley Department of Electrical Engineering & Data Science Institute Columbia University

2 PRINCIPAL COMPONENT ANALYSIS

3 DIMENSIONALITY REDUCTION We re given data x 1,..., x n, where x R d. This data is often high-dimensional, but the information doesn t use the full d dimensions. For example, we could represent the above images with three numbers since they have three degrees of freedom. Two for shifts and a third for rotation. Principal component analysis can be thought of as a way of automatically mapping data x i into some new low-dimensional coordinate system. It capture most of the information in the data in a few dimensions Extensions allow us to handle missing data, and unwrap the data.

4 PRINCIPAL COMPONENT ANALYSIS Example: How can we approximate this data using a unit-length vector q? x i q is a unit-length vector, q T q = 1. q (q T x i )q Red dot: The length, q T x i, to the axis after projecting x onto the line defined by q. The vector (q T x i )q takes q and stretches it to the corresponding red dot. So what s a good q? How about minimizing the squared approximation error, q = arg min q x i qq T x i 2 subject to q T q = 1 qq T x i = (q T x i )q : The approximation of x i by stretching q to the red dot.

5 PCA : THE FIRST PRINCIPAL COMPONENT This is related to the problem of finding the largest eigenvalue, q = arg min q = arg min q x i qq T x i 2 s.t. q T q = 1 ( ) xi T x i q T x i xi T q }{{} = XX T We ve defined X = [x 1,..., x n ]. Since the first term doesn t depend on q and we have a negative sign in front of the second term, equivalently we solve q = arg max q q T (XX T )q subject to q T q = 1 This is the eigendecomposition problem: q is the first eigenvector of XX T λ = q T (XX T )q is the first eigenvalue

6 PCA: GENERAL The general form of PCA considers K eigenvectors, q = arg min q = arg min q x i xi T x i K { (xi T q k )q k 2 s.t. q T 1, k = k k q k = 0, k k k=1 }{{} approximates x K k=1 q T k ( ) x i xi T } {{ } = XX T q k The vectors in Q = [q 1,..., q K ] give us a K dimensional subspace with which to represent the data: q T 1 x K x proj =., x (q T k x)q k = Qx proj q T k=1 Kx The eigenvectors of (XX T ) can be learned using built-in software.

7 EIGENVALUES, EIGENVECTORS AND THE SVD An equivalent formulation of the problem is to find (λ, q) such that (XX T )q = λq Since (XX T ) is a PSD matrix, there are r min{d, n} pairs, λ 1 λ 2 λ r > 0, q T k q k = 1, q T k q k = 0 Why is (XX T ) PSD? Using the SVD, X = USV T, we have that (XX T ) = US 2 U T Q = U, λ i = (S 2 ) ii 0 Preprocessing: Usually we first subtract off the mean of each dimension of x.

8 PCA: EXAMPLE OF PROJECTING FROM R 3 TO R 2 For this data, most information (structure in the data) can be captured in R 2. (left) The original data in R 3. The hyperplane is defined by q 1 and q 2. [ ] (right) The new coordinates for the data: x i x proj x T i = i q 1 xi T. q 2

9 EXAMPLE: DIGITS Data: images of handwritten 3 s (as vectors in R 256 ) Mean λ 1 = λ 2 = λ 3 = λ 4 = Above: The first four eigenvectors q and their eigenvalues λ. Original M = 1 M = 10 M = 50 M = 250 Above: Reconstructing a 3 using the first M 1 eigenvectors plus the mean, and approximation M 1 x mean + (x T q k )q k k=1

10 PROBABILISTIC PCA

11 PCA AND THE SVD We ve discussed how any matrix X has a singular value decomposition, X = USV T, U T U = I, V T V = I and S is a diagonal matrix with non-negative entries. Therefore, XX T = US 2 U T (XX T )U = US 2 U is a matrix of eigenvectors, and S 2 is a diagonal matrix of eigenvalues.

12 A MODELING APPROACH TO PCA Using the SVD perspective of PCA, we can also derive a probabilistic model for the problem and use the EM algorithm to learn it. This model will have the advantages of: Handling the problem of missing data Allowing us to learn additional parameters such as noise Provide a framework that could be extended to more complex models Gives distributions used to characterize uncertainty in predictions etc.

13 PROBABILISTIC PCA In effect, this is a new matrix factorization model. With the SVD, we had X = USV T. We now approximate X WZ, where W is a d K matrix. In different settings this is called a factor loadings matrix, or a dictionary. It s like the eigenvectors, but no orthonormality. The ith column of Z is called zi R K. Think of it as a low-dimensional representation of x i. The generative process of Probabilistic PCA is x i N(Wz i, σ 2 I), z i N(0, I). In this case, we don t know W or any of the z i.

14 THE LIKELIHOOD Maximum likelihood Our goal is to find the maximum likelihood solution of the matrix W under the marginal distribution, i.e., with the z i vectors integrated out, W ML = arg max W ln p(x 1,..., x n W) = arg max W ln p(x i W). This is intractable because p(x i W) = N(x i 0, σ 2 I + WW T ), N(x i 0, σ 2 I + WW T 1 ) = (2π) d 2 σ 2 I + WW T 1 2 e 1 2 xt (σ 2 I+WW T ) 1 x We can set up an EM algorithm that uses the vectors z 1,..., z n.

15 EM FOR PROBABILISTIC PCA Setup The marginal log likelihood can be expressed using EM as ln p(x i, z i W) dz i = + q(z i ) ln p(x i, z i W) q(z i ) q(z i ) ln dz i L q(z i ) p(z i x i, W) dz i KL EM Algorithm: Remember that EM has two iterated steps 1. Set q(z i ) = p(z i x i, W) for each i (making KL = 0) and calculate L 2. Maximize L with respect to W Again, for this to work well we need that we can calculate the posterior distribution p(z i x i, W), and maximizing L is easy, i.e., we update W using a simple equation

16 THE ALGORITHM EM for Probabilistic PCA Given: Data x 1:n, x i R d and model x i N(Wz i, σ 2 ), z i N(0, I), z R K Output: Point estimate of W and posterior distribution on each z i E-Step: Set each q(z i ) = p(z i x i, W) = N(z i µ i, Σ i ) where Σ i = (I + W T W/σ 2 ) 1, µ i = Σ i W T x i /σ 2 M-Step: Update W by maximizing the objective L from the E-step [ ] [ W = x i µ T i σ 2 I + (µ i µ T i + Σ i ) ] 1 Iterate E and M steps until increase in n ln p(x i W) is small. Comment: The probabilistic framework gives a way to learn K and σ 2 as well.

17 EXAMPLE: IMAGE PROCESSING = 8 x 8 patch X data matrix, e.g., 64 x 262,144 For image problems such as denoising or inpainting (missing data) Extract overlapping patches (e.g., 8 8) and vectorize to construct X Model with a factor model such as Probabilistic PCA Approximate x i Wµ i, where µ i is the posterior mean of z i Reconstruct the image by replacing x i with Wµ i (and averaging)

18 EXAMPLE: DENOISING Noisy image on left, denoised image on right. The noise variance parameter σ 2 was learned for this example.

19 E XAMPLE : M ISSING DATA Another somewhat extreme example: I Image is (RGB dimension) I Throw away 80% at random I (left) Missing data, (middle) reconstruction, (right) original image

20 KERNEL PCA

21 KERNEL PCA We ve seen how we can take an algorithm that uses dot products, x T x, and generalize with a nonlinear kernel. This generalization can be made to PCA. Recall: With PCA we find the eigenvectors of the matrix n x ix T i = XX T. Let φ(x) be a feature mapping from R d to R D, where D d We want to solve the eigendecomposition [ ] φ(x i )φ(x i ) T q k = λ k q k without having to work in the higher dimensional space. That is, how can we do PCA without explicitly using φ( ) and q?

22 KERNEL PCA Notice that we can reorganize the operations of the eigendecomposition φ(x i ) ( φ(x i ) T ) q k /λk = q k }{{} = a ki That is, the eigenvector q k = n a kiφ(x i ) for some vector a k R n. The trick is that instead of learning q k, we ll learn a k. Plug this equation for q k back into the first equation: N φ(x i ) j=1 a kj φ(x i ) T φ(x j ) }{{} = K(x i,x j) = λ k a ki φ(x i ) and multiply both sides by φ(x l ) T for each l {1,..., n}.

23 KERNEL PCA When we multiply the following by φ(x l ) T for l = 1..., n: N φ(x i ) j=1 we get a new set of linear equations a kj φ(x i ) T φ(x j ) }{{} = K(x i,x j) = λ k a ki φ(x i ) K 2 a k = λ k Ka k Ka k = λ k a k where K is the n n kernel matrix constructed on the data. Because K is guaranteed to be PSD because it is a matrix of dot-products, the LHS and RHS above share a solution for (λ k, a k ). Now perform regular PCA, but on the kernel matrix K instead of the data matrix XX T. We summarize the algorithm on the following slide.

24 KERNEL PCA ALGORITHM Kernel PCA Given: Data x 1,..., x n, x R d, and a kernel function K(x i, x j ). Construct: The kernel matrix on the data, e.g., K ij = b exp Solve: The eigendecomposition { xi xj 2 c }. Ka k = λ k a k for the first r n eigenvector/eigenvalue pairs (λ 1, a 1 ),..., (λ r, a r ). Output: A new coordinate system for x i by (implicitly) mapping φ(x i ) and then projecting q T k φ(x i) x i projection λ 1 a 1i. λ r a ri where a ki is the ith dimension of the kth eigenvector a k.

25 KERNEL PCA AND NEW DATA Q: How do we handle new data, x 0? Before, we could take the eigenvectors q k and project x T 0 q k, but a k is different here. A: Recall the relationship of a k to q k in kernel PCA is q k = a ki φ(x i ). We used the kernel trick to avoid working with or even defining φ(x i ). As with regular PCA, after mapping x 0 we want to project onto eigenvectors x 0 projection φ(x 0 ) T q 1. φ(x 0 ) T q r Plugging in for q k : φ(x 0 ) T q k = a ki φ(x 0 ) T φ(x i ) = a ki K(x 0, x i ).

26 EXAMPLE RESULTS An example of kernel PCA using the Gaussian kernel. (left) Original data, colored for reference (but may be classes) (middle) New coordinates using kernel width c = 2 (right) New coordinates using kernel width c = 10 Terminology: What we are doing is closely related to spectral clustering and can be considered an instance of manifold learning.