Download presentation
Presentation is loading. Please wait.
1
Unsupervised Learning - PCA The neural approach->PCA; SVD; kernel PCA Hertz chapter 8 Presentation based on Touretzky + various additions
28
Use a set of pictures of faces to construct a PCA space. Shown are first 25 principal components. (C. DeCoro)
30
Variance as function of number of PC
31
Each iteration, from left to right, corresponds to addition of 8 principal components
32
Matching general images to faces using eigenfaces space
34
Singular Value Decomposition Singular Value Decomposition This can be rewritten in the matrix representation where ∑is a (non-square) diagonal matrix, and U,V are orthogonal matrices. Ordering the non-zero elements of ∑ in descending order, we can get an approximation of lower rank r of the matrix X by choosing =0 for j>r leading to SVD involves expanding an mxn matrix X of rank k=min(m,n) into a sum of k unitary matrices of rank 1, in the following way:
35
Singular Value Decomposition continued This is the best approximation of rank r to X, i.e. it leads to the minimal sum of square deviations
36
Relation between SVD and PCA SVD uses the same unitary transformations as PCA performed on the rows or columns of X (using the columns or rows as feature spaces). The singular values of SVD are the square toots of the eigenvalues of PCA.
37
SVD – Singular Value Decomposition
38
Applications of SVD Dimensionality reduction, compression Noise reduction Pattern search, clustering Example: microarray of expression data, “ DNA chips ”
39
Affymetrix GeneChip ® probe array Image courtesy of Affymetrix
40
Hybridization of tagged probes Image courtesy of Affymetrix
41
Microarray Experiment Result
42
SVD example: gene expressions of nine rats
43
SVD – Singular Value Decomposition
44
Alter et al 2000: identify eigengenes that are responsible for yeast cell cycle
45
Processing: SVD
46
Processing: SVD – 1 dimension
47
Processing: SVD – 10 dimensions
48
Processing: SVD – 100 dimension
49
Kernel PCA
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.