LING 696B: PCA and other linear projection methods

1

LING 696B: PCA and other linear projection methods

2

Curse of dimensionality The higher the dimension, the more data

is needed to draw any conclusion Probability density estimation:

Continuous: histograms

Discrete: k-factorial designs Decision rules:

Nearest-neighbor and K-nearest neighbor

3

How to reduce dimension? Assume we know something about

the distribution Parametric approach: assume data follow

distributions within a family H Example: counting histograms for 10-

D data needs lots of bins, but knowing it’s a pancake allows us to fit a Gaussian (Number of bins)10 v.s. (10 + 10*11/2)

4

Linear dimension reduction Pancake/Gaussian assumption is

crucial for linear methods Examples:

Principle Components Analysis Multidimensional Scaling Factor Analysis

5

Covariance structure of multivariate Gaussian 2-dimensional example

No correlations --> diagonal covariance matrix, e.g. Special case: = I - log likelihood Euclidean distance to the

center

Variance in each dimension

Correlation between dimensions

6

Covariance structure of multivariate Gaussian Non-zero correlations --> full

covariance matrix, COV(X1,X2) 0 E.g. =

Nice property of Gaussians: closed under linear transformation

This means we can remove correlation by rotation

7

Covariance structure of multivariate Gaussian Rotation matrix: R = (w1, w2),

where w1, w2 are two unit vectors perpendicular to each other Rotation by 90 degree

Rotation by 45 degree

w1 w2

w1

w2

w1w2

8

Covariance structure of multivariate Gaussian Matrix diagonalization: any 2X2

covariance matrix A can be written as:

Interpretation: we can always find a rotation to make the covariance look “nice” -- no correlation between dimensions

This IS PCA when applied to N dimensions

Rotation!

9

Computation of PCA The new coordinates uniquely identify

the rotation

In computation, it’s easier to identify one coordinate at a time.

Step 1: centering the data X <-- X - mean(X) Want to rotate around the center

w1w2

w3

3-D: 3 coordinates

10

Computation of PCA Step 2: finding a direction of

projection that has the maximal “stretch”

Linear projection of X onto vector w: Projw(X) = XNXd * wdX1 (X centered)

Now measure the stretch This is sample variance = Var(X*w)

wx X

w

11

Computation of PCA Step 3: formulate this as a

constrained optimization problem Objective of optimization: Var(X*w) Need constraint on w: (otherwise can

explode), only consider the direction So formally:

max||w||=1 Var(X*w), find w

12

Computation of PCA Some algebra (homework):

Var(x) = E[(x - E[x])2

= E[x2] - (E[x])2

Apply to matrices (homework)Var(X*w) = wT XT X w = wTCov(X) w (why)

Cov(X) is a dXd matrix (homework) Symmetric (easy) For any y, yTCov(X) y >= 0 (tricky)

13

Computation of PCA Going back to the optimization

problem:= max||w||=1 Var(X*w)= max||w||=1 wTCOV(X) w

The answer is the largest eigenvalue for COV(X)

w1

The first Principle Component!

(see demo)

14

More principle components We keep looking among all the

projections perpendicular to w1

Formally:max||w2||=1,w2w1 wTCov(X) w

This turns out to be another eigenvector corresponding to the 2nd largest eigenvalue(see demo) w2

New coordinates!

15

Rotation Can keep going until we find all

projections/coordinates w1,w2,…,wd

Putting them together, we have a big matrix W=(w1,w2,…,wd)

W is called an orthogonal matrix This corresponds to a rotation

(sometimes plus reflection) of the pancake

This pancake has no correlation between dimensions (see demo)

16

When does dimension reduction occur? Decomposition of covariance

matrix

If only the first few ones are significant, we can ignore the rest, e.g. 2-D coordinates of X

17

Measuring “degree” of reduction

a2a1

Pancake data in 3D

18

Reconstruction from principle components Perfect reconstruction (x

centered):

Reconstruction error:

w1w2

xlength

direction

Another fomulationof PCA

x

Many pieces

The bigger pieces

19

A creative interpretation/ implementation of PCA Any x can be reconstructed from

principle components (PC form a basis for the whole space)

Output X

Input X

hidden=W

W

When (# of hidden) < (# of input), the network does dimension reduction

This can be used to implement PCA

“neural firing” Connection weights

“encoding”

20

An intuitive application of PCA:(Story and Titze) and others Vocal tract measurements are high

dimensional (different articulators) Measurements from different positions are

correlated Underlying geometry: a few articulatory

parameters, possibly pancake-like after collapsing a number of different sounds

Big question: relate low-dimensional articulatory parameters (tongue shape) to low dimensional acoustic parameters (F1/F2)

21

Story and Titze’s application of PCA Source data: area function data

obtained from MRI (d=44) Step 1: Calculate the mean

Interestingly, the mean produces a schwa-like frequency response

22

Story and Titze’s application of PCA Step 2: substract the mean from

the area function (center the data)

Step 3: form the covariance matrix

R = XTX (dXd matrix), X ~ (x, p)

23

Story and Titze’s application of PCA

Step 4: eigen-decomposition of the covariance matrix, get PC’s Story calls them “empirical modes”

Length of projection:

Reconstruction:

24

Story and Titze’s application of PCA

Story’s principle components The first 2 PC’s can do most of the

reconstruction Can be seen as a perturbation of overall

tongue shape (from the mean)

Constriction < 0

Expansion > 0

25

Story and Titze’s application of PCA The principle components are

interpretable as control parameters

Acoustic-to-Articulatory mapping almost one-to-one after dimension reduction

26

Applying PCA to ultrasound data? Another imaging technique

Generate a tongue profile similar to X-ray and MRI

High-dimensional Correlated Need dimension reduction to interpret

articulatory parameters See demo

27

An unintuitive application of PCA Latent Semantic Indexing in

document retrieval Documents as vectors of word counts Try to extract some “features” by

linear combination of word counts The underlying geometry unclear

(mean? Distance?) The meaning of principle components

unclear (rotation?)

“market”

“stock”

“bonds”

28

Summary of PCA: PCA looks for:

A sequence of linear, orthogonal projections that reveal interesting structure in data (rotation)

Defining “interesting”: Maximal variance under each

projection Uncorrelated structure after

projection

29

Departure from PCA 3 directions of divergence

Other definitions of “interesting”? Linear Discriminant Analysis Independent Component Analysis

Other methods of projection? Linear but not orthogonal: sparse coding Implicit, non-linear mapping

Turning PCA into a generative model Factor Analysis

30

Re-thinking “interestingness” It all depends on what you want Linear Disciminant Analysis (LDA):

supervised learning Example: separating 2 classes

Maximal variance

Maximal separation

31

Re-thinking “interestingness” Most high-dimensional data look like

Gaussian under linear projections Maybe non-Gaussian is more interesting

Independent Component Analysis Projection pursuits

Example: ICA projection of 2-class dataMost unlike Gaussian (e.g. maximize kurtosis)

32

The “efficient coding” perspective

Sparse coding: Projections do not have to be orthogonal There can be more basis vectors than

the dimension of the space Neural interpretation (Dana Ballard’s talk last

week)xw2

w1

w3

w4

p << d; compact coding (PCA)p > d; sparse coding

Basis expansion

33

“Interesting” can be expensive

Often faces difficult optimization problems Need many constraints Lots of parameter sharing Expensive to compute, no longer an

eigenvalue problem

34

PCA’s relatives: Factor Analysis PCA is not a generative model:

reconstruction error is not likelihood Sensitive to outliers Hard to build into bigger models

Factor Analysis: adding a measurement noise to account for variability

Factors: spherical Gaussian N(0,I)

observation

Loading matrix (scaled PC’s)

Measurement noiseN(0,R), R diagonal

35

PCA’s relatives: Factor Analysis Generative view: sphere --> stretch

and rotate --> add noise

Learning: a version of EM algorithm (see demo and synthesis)

36

Mixture of Factor Analyzers Same intuition as other mixture

models: there may be several pancakes out there, each with its own center/rotation

37

PCA’s relatives: Metric multidimensional scaling Approach the problem in a different

way No measurements from stimuli, but

pairwise “distance” between stimuli Intend to recover some

psychological space for the stimuli See Jeff’s talk