You are on page 1of 21

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

NSF-CBMS Lecture Series


Lecture 1:
Finding Hidden Components
By Data Analysis

Agenda

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

Hidden Components
Principal Components
Beyond Principal Components

David Donoho
Statistics Department
Stanford University

Statistical Analysis of Images

% &

Look Ahead

&

From Newtons Notebook

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

(b) Prism

% &

'

(a) Newton

I.A Hidden Components in Light

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

&

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

I.B Hidden Components in Heat

Figure 2: Fourier

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

Diffusion of Heat

f (, t), for [0, 2)

f (, t = 0), for [0, 2)

Initial Temperature Profile:

Profile at later time:

Shows
less variation
more smoothness
eventual uniformity

Stages of Diffusion

Fouriers Insight

cn exp{in}

n (t) cn exp{in}

Expand in Sinusoidal Components


X

f (, t = 0) =

Profile made of waves.

f (, t) =

Profile at later time:

Damping factor:

n (t) = exp{n2 t}

Decreases rapidly with wavelength, time

% &

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

&

% &

$ '

'

1.5

0.5

-0.5

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

&

f(,t)

'

1.5
1
0.5
0
-0.5
0

Initial Density t=0

Components t=0

10

Components t=.0469

f(,t=.)

10

Component

Diffusion

Fourier1

Fourier

1.5

(cn (t))

(cn (0))

Modern Viewpoint I
f (, 0)

f (, t)

0.5

-0.5

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

&

'

Component

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

&

f(,0)

Density at t=.0469

n (t)

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

Interpretation

Fourier Analysis is a kind of Prism

Hidden Components: Sinusoidal Components

Post Facto: Sinusoids Fundamental

Dominant trend, 19th century Mathematical Physics

10

Tn,m cm

12

% &

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

11

$ '

nm

Modern Viewpoint II
Linear Transformation
(T c)n =

Diagonal Transformation

(T c)n = Tn,n cn

Diagonal no interaction between components.

% &

'

f (, 0)
Fourier1

Fourier

(cn (t))

(cn (0))

Modern Viewpoint III

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Diffusion
f (, t)

Diagonal

13

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

II. Principal Components

14

(b) Lo`eve

% &

(a) Karhunen

&

hn , m i = n,m

(t, s)n (s)ds = n n (t)

1 2 3 . . .

16

Eigenvectors

Under general conditions, orthobasis 1 , 2 , . . .,

Cov(X(t), X(s)) = E(X(t) EX(t))(X(s) EX(s))

(t, s) its covariance:

X(t), t T a random process in L2 (T )

Karhunen-Lo`
eve Transform

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

(b) Hotelling

Eigenvalues
% &

15

'

(a) Pearson

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

&

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

f
KLT1

KLT

(dn )

(cn )

Modern Viewpoint

Covariance
f

17

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

KLT Representation

X(t), t T a random process in L2 (T )

Xn n (t)

Cov(Xn , Xm ) = n n,m

V arXn = n

Xn = hX, n i, its n-th coordinate in KL basis


Variances
Covariances
Representation
X(t) =

18

% &

here the Xn are orthogonal.


&

20

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

Noise Removal
Y (t) observed data
X(t) desired signal to recover

Y (t) = X(t) + N (t)

wn Yn n (t)

n
.
n + n

N (t) Gaussian noise process, with


Same KL basis as X: (n )

wn =

E{X(t)|Y } =

Eigenvalues (n )

where

% &

19

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Some Key Applications


Noise Removal Wiener
Data Compression Shannon

&

'

KLT

Modern Viewpoint
Y

KLT1

n)
(X

(Yn )

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Expectation
E{X(t)|Y }

wn

21

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

Data Compression

22

0.1

0.2

0.3

0.4

Brownian Bridge

0.5
t

0.6

0.7

24

% &

30

25

20

15

10

-5

-10

-15

-20

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

% &

23

&

0.9

$ '

0.8

'

Key Example

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

B(t) Brownian Bridge.


Gaussian
B(0) = B(1) = 0
(t, s) = min(t, s) ts.
Closely related to Brownian Motion

&

B(t)

'

2 sin(nt)

X
Zn n (t)

KLT of Brownian

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

n (t) =
n (0) = n (1) = 0
n = Const/n2
Representation:
B(t) =
Principal Scores Zn indep N (0, n ).

25

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

III. Beyond Principal Components


Independent Components

26

Sparse Components

% &

Sinusoids are the Principal Components of Brownian Bridge.


&

Independent Components of Gaussian Process


For Gaussian Process, IC = PC.

28

n n (t)

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

X
n

Proof:

If (Xn ) uncorrelated & jointly Gaussian, the (Xn ) are


independent.

For a Gaussian Process, the principal scores are uncorrelated


and jointly Gaussian.

% &

27

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

X(t) =

III.A Independent Components

n Independent
n Basis Functions
Remarks
Need not exist
Need not be orthogonal
No procedure in general
&

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

t
t1
t

t<

Ramp: A NonGaussian Process


Scalar RV U (0, 1).

X(t, ) =

Stochastic Process (Y. Meyer)

Properties:
EX(t) = 0

29

0.1

0.2

0.3

0.4

Ramp

0.5
t

0.6

0.7

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '
1

0.8

0.6

0.4

0.2

-0.2

0.8

0.9

30

-0.4

% &

X(1) = X(0) = 0
&

0.1

0.1

0.2

0.2

0.3

0.3

0.4

0.4

Brownian Bridge

0.5
t
Ramp

0.5
t

0.6

0.7

0.7

0.8

0.8

0.9

0.9

32

$
50
40
30
20

10

-10
-20

1
0.8
0.6
0.4

0.2

-0.2
-0.4

0.6

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

% &

31

$ '

&

Second-Order Equivalent to Brownian Bridge!

(t, s) = Cov(X(t), X(s)) = min(t, s) ts

Covariance of Ramp

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

'

Ramp(t)
B(t)
Ramp(t)

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Contrast
Principal Components are second-order independent, but
perhaps not independent.

33

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

III.B Approximate ICA

Attempt to approximate Higher-Order Independence


Form Cumulant Tensor of order > 2

34

Attempt to Diagonalize Tensor

% &

Independent Components are independent, but perhaps cannot


be found or do not exist.

&

E(X1 X2 ) E(X3 X4 )

E(X3 ) E(X1 X2 X4 ) E(X4 ) E(X1 X2 X3 )

E(X1 ) E(X2 X3 X4 ) E(X2 ) E(X1 X3 X4 )

E(X1 X2 X3 X4 )

Independence and Cumulant Tensors

Let Xn = hX, n i coefficient in basis (n ).


Cumulant Tensor K4, 4-way array:

K4, (i1 , i2 , i3 , i4 ) = Cum4 (Xi1 , Xi2 , Xi3 , Xi4 )

If (X ) are independent random variables, K4, is hyperdiagonal:


n

0 not all indices same


? i 1 = i2 = i 3 = i4
K4, (i1 , i2 , i3 , i4 ) =

36

E(X1 X3 ) E(X2 X4 ) + E(X1 X4 ) E(X2 X3 )

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

2(EX1 )(EX2 )(EX3 )(EX4 )

% &

35

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Fourth-order Joint Cumulants

Kurtosis: K4 (X) = E(X )4 3V ar(X)2

Cum4 (X1 , . . . , X4 )

&

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Numerical Experiment
Finite Ramp Process:
T = {1/32, . . . , 32/32}
= {1/32, . . . , 32/32}
Database X(i, j) =
X(i , tj ).

N = 32 realizations X(i, )
N = 32 variables X(, t)
Use Jean-Francois Cardoso JADE Algorithm
Attempt Diagonalization of Empirical Cumulant Tensor
Moves 90% of Kurtic Energy on Diagonal.

37

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

20

20

20

20

20

Results from JADE analysis of Ramp

-1
10

20

-1
10

20

-1
10

-1
10

-1
10

-1
10

-1
10

20

20

20

-1
10

-1

20

20

20

20

20

20

20

0
-1
10
0

-1
10
0
-1
10
0
-1
10
0
-1
10
0
-1
10
0

20

0
-1

-1
10

20

20

20

20

20

20

20

-1
10
0
-1
10
0
-1
10
0
-1
10
0
-1
10
0

20

0
-1

-1
10

20

20

20

20

20

20

0
-1
10
0
-1
10
0
-1
10
0
-1
10
0
-1
10
0

-1
10

-1
10
0
-1

Jadelet 16

Jadelet 9

Comparisons
Jadelet 5

0.5

0.5

0.5

-1

0.5

20

-1

20

-0.5

-1

Wavelet 23

-0.5

20

0.5

Wavelet 21

-0.5

-1

Wavelet 12

0.5

0.5

-1

0.5

20

-1

20

-0.5

-1

-0.5

20

-0.5

-1

-0.5

-0.5

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Wavelet 25

20

38

40

%
39

% &

20

20

20

20

20

20

20

20

Jadelet 13

&

20

0
-1
10
0

-1

$ '

20

% &

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

20

-1
10

20

0
-1
10

20
0
-1
10
0

20

0
-1
10
0
-1
10

20

20

20

0
-1
10
0
-1
10
0

-1
10

20

-1
10

-1
10

-1
10

-1
10

20

20

20

20

-1
10

-1
10

-1
10

-1
10

20

20

20

20

-1
10

-1
10

-1
10

-1
10

20

20

20

20

-1
10

-1

0
20

-1
10

0
-1

-1
10

20

0
0

-1
10

-1

-1
10

-1
10

-1
10

20

Daubechies Nearly Symmetric Wavelets

&

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Interpretations
Hidden Components discovered by JADE
Dyadic Scales
Near-Translates
Approx. Daubechies Nearly Symmetric

41

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

Xn n (t)

III.D Sparse Components Analysis


X(t) =

(Xn ) a sparse vector: few large entries.

n a basis or dictionary to be found

Goal choose (n ) to optimize sparsity


Remarks:
Need not exist
No general definition

42

Quasi-Wavelets are the almost-independent components of Ramp

% &

No general algorithm
&

cnj nj k2

44

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

j=1

m
X

Comparison of SCA with PCA (cont)


SCA (One approach)

B(m) (X, (n )) best m-term approximation,

kX B(m) (X, (n ))k2 = argmin(nj ) kX

argmin(n ) EkX B(m) (X, (n ))k22

Find Basis optimizing

% &

43

'

Comparison of SCA with PCA

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

PCA

m
X
n=1

cn n

A(m) (X, (n )) first m-term approximation,


A(m) X =

argmin(n ) EkX A(m) (X, (n ))k22

Find Basis (n ) optimizing

&

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

X
j,k j,k (t)

Example of Approximate SCA: Ramp


Wavelet Representation
X(t) =
n

(j,k ) Orthonormal Wavelets basis, 3 vanishing Moments.


(j,k ) sparse vector: few large entries.
At level j at most C nonzeros, k
With 54 coefficients can represent to accuracy 106

160

45

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

Sparsity Comparison
Wavelet Basis (quasi-SCA)

With 54 coefficients can represent to accuracy 106

kX B(m) (X, wavelets)k2 C1 exp(C2 m)


Fourier Basis (PCA)

Need 1902 coefficients for accuracy 106 .

kX A(m) (X, sinusoids)k2 C/ m

kX B(m) (X, sinusoids)k2 C/ m

1
0.8
0.6
0.4

0.2

-0.2
-0.4

-2
-4
-6
-8
-10

0.2

0.6

Ramp

0.4

0.4

0.6

0.8

0.8

Wavelet Transform of Ramp

0.2

Coefficient Amplitude

0.5

-0.5

10

-2

10

-4

10

-6

10

200

100
200
k-th largest coefficient

Wavelet Transform

Sine Transform

Ordered Transform Coefficients

100

Sine Transform of Ramp

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

46

48

140

47

% &

300

&

120

Compression Numbers for DST and Wavelet Analysis of Ramp

80

Sine Transform

60

Wavelet Transform

40

300

200

$ '

20

180

'

-1

-2

-3

-7

-6

-5

-4

-8

100
n

% &

Ramp(t)

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

&

Log(c )

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Second-Order

Independence

High-Order

ICA

First-m Term

Approximation

Best-m Term

SCA

Contrast of ICA/SCA/PCA

New Concept

PCA Concept

49

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Aspect

Immense

ICA

Approximation Theory

Rarified

SCA

$ '

Popularity

Statistical

Comparison of ICA/SCA

Mode

Plausibility

Many Heuristics

None

Few Heuristics

Sufficient

Harmonic Analysis

Algorithms

50

Approximation

% &

Independence

&

|K4, (i1 , i2 , i3 , i4 )|p )1/p

52

i1 ,i2 ,i3

Almost-Diagonality of Cumulant Tensors

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

51

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Similar Outcome from ICA & SCA: Ramp

i1

Interesting case; p 0; says each slice sparse.

kK4, kAD(p) = sup(

Wavelets give very high rate of m-term Approximation.

Theorem. For circular Ramp process on [0, 2), Daubechies nearly


symmetric wavelets with 3 vanishing moments

% &

Almost-diagonal in sense of good decay away from hyper-diagonal.

kK4, kAD(p) <

Wavelets almost diagonalize the cumulant tensor.


Can find qualitatively same outcome by either principle.

&

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Interpretation
Study 1 d objects with singularities
They are nongaussian
PCA would give sinsuoidal components
ICA & SCA give wavelet components

53

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

Retina

...

V1

IV. Visual Neuroscience

Information Flow in the Visual System


Image

V1 Component of Visual Cortex:

Linear Receptive Fields Hubel and Wiesel.

Neural Outputn = hImage, n i


Properties of (n )
Localized
Oriented

...

54

What if we study 2 d objects Images?

% &

Wavelet/Gabor Models
&

(b) Field

56

(a) Barlow

IV.A Two Slogans

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

% &

55

'

Caricature of V1

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

&

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Barlow
Visual System does a remarkable job of image compression.
107 bits/sec at input
20 40 bits/sec at end of processing

57

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

Field

Structure of natural images drives structure of visual system

Redundancy of natural images could teach visual system how


to be sparse

58

Statistical analysis of images and neurodevelopment? (Cats)

% &

Visual system takes highly redundant data, removes the


redundancy, produces nearly independent data.

&

60

Depict results

Sparse Components Analysis

Database of patches

Extract 16 16 patches.

Collection of natural images

Appeared in Nature, 1996

Olshausen-Field Experiment

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

(b) Field

% &

59

'

(a) Olshausen

Sparse Components of Natural Images

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

&

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Database of Image Patches


Collection of natural images
Extract 16 16 patches.
Patch P (p) = (P (p) (s, t), 1 s, t 16
X(p, ) = entries of P (p) in row-major order.
Result: X an N by 256 matrix

61

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

j (p)j

Empirical Sparse Components Analysis


Want X(p, j)

kX(p, )

Xj (p)j k2 + S((p))

Find both coefficients j (p) and basis j .


Want j (p) sparse.

Formal Objective
X
(j ) (j (p))

min min

Sparsity measure

log(1 + j2 )

62

% &

S() =
&

64

Interpretation

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

Multi-Oriented
Multiscale
Bandpass
Initially, wavelets. Later not in title.
Clearly (to me) not wavelets.

% &

63

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

&

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

66

65

$ '

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

% &

Beamlets

Curvelets

Ridgelets

Similarities to Recent CHA ideas

Not Gabor too many scales

Not Wavelets too many directions

What Are These Things?

&

68

van der Schaaf and van Hateren

Hyvarinen

Lewicki-Sejnowksi

Hundreds of papers

NIPS Conferences

ICA-2000 Conference

Independent Components Analysis

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

(b) Sejnowksi

% &

67

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

(a) Bell

Independent Components of Natural Images

&

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Independent Components of Natural Movies

69

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

A Role for Math?

PCA Data Analysis finds Sinusoids

ICA/SCA Data Analysis of 1-d Discontinuities finds Wavelets

ICA/SCA Data Analysis of 2-d Discontinuities finds ???

ICA/SCA Data Analysis of 3-d Discontinuities finds ???

70

Little Prospect of substantially larger data analysis

% &

Figure 8: Hans van Hateren

&

Wed.

Curvelets

Thurs.

Combined

Beamlets

Fri.

72

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

Tues.

Ridgelets

Curvelets

Lecture Plan
Mon.

Ridgelets

Curvelets

71

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Lectures to Come
Analogy between Statistical Analysis of Image Databases&
Harmonic Analysis of Image Models.

Data Anal.

Ridgelets

Common Factor: Sparse Representation


Point Singularities vs. Line Singularities

Math Anal.

% &

Line Singularities vs. Curve Singularities


Harmonic Analysis of Line & Curve Singularities
What Lies Beyond
&

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Whats The Connection?

73

Whats at Stake

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

Applications
Data Compression
Noise Removal
Detection, Classification
New Approximation Theory

New Connections with Mathematics

New Research in Harmonic Analysis

74

(b) Imagelets

% &

(a) Ridgelets

&

(b) X. Huo

76

(a) E. Cand`es

Collaborators (2)

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

% &

75

'

Figure 9: Iain Johnstone

Collaborators (1)

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

&

'

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Collaborators (3ab)

(b) Coifman

77

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

$ '

(b)
Wald
en

Collaborators (3cd)

(a) Israeli

78

(a) Averbuch

% &

(b) Y. Hel-Or

80

&

(a) H. Hel-Or

Collaborators (5)

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

Collaborators (4)

(b) E. Arias

% &

79

$ '

(c) O. Levi

'

(a) G. Flesia

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

&

'

(a) Stodden

Thanks ...

(b) Duncan

NSF-CBMS Talk 1: Finding Hidden Components by Data Analysis

&

81

You might also like