Pattern Recognition

advertisement
Spoken Dialog Systems and Voice
XML :
Intro to Pattern Recognition
Esther Levin
Dept of Computer Science
CCNY
Some materials used in this course were taken from the textbook “Pattern Classification” by Duda et al., John Wiley & Sons, 2001
with the permission of the authors and the publisher
Credits and Acknowledgments
Materials used in this course were taken from the textbook “Pattern
Classification” by Duda et al., John Wiley & Sons, 2001 with the permission of
the authors and the publisher; and also from
Other material on the web:












Dr. A. Aydin Atalan, Middle East Technical University, Turkey
Dr. Djamel Bouchaffra, Oakland University
Dr. Adam Krzyzak, Concordia University
Dr. Joseph Picone, Mississippi State University
Dr. Robi Polikar, Rowan University
Dr. Stefan A. Robila, University of New Orleans
Dr. Sargur N. Srihari, State University of New York at Buffalo
David G. Stork, Stanford University
Dr. Godfried Toussaint, McGill University
Dr. Chris Wyatt, Virginia Tech
Dr. Alan L. Yuille, University of California, Los Angeles
Dr. Song-Chun Zhu, University of California, Los Angeles
Outline
Introduction

What is this pattern recogntiion
Background Material

Probability theory
PATTERN RECOGNITION AREAS
Optical Character Recognition ( OCR)


Sorting letters by postal code.
Reconstructing text from printed materials (such as reading machines for blind
people).
Analysis and identification of human patterns


Speech and voice recognition.
Finger prints and DNA mapping.
Banking and insurance applications


Credit cards applicants classified by income, credit worthiness, mortgage amount, # of
dependents, etc.
Car insurance (pattern including make of car, #of accidents, age, sex, driving habits,
location, etc).
Diagnosis systems


Medical diagnosis (disease vs. symptoms classification, X-Ray, EKG and tests
analysis, etc).
Diagnosis of automotive malfunctioning
Prediction systems
Weather forecasting (based on satellite data).
 Analysis of seismic patterns
Dating services (where pattern includes age, sex, race, hobbies, income, etc).

More Pattern Recognition
Applications
SENSORY
Vision

Face/Handwriting/Hand
Speech

Speaker/Speech
Olfaction

Apple Ripe?
DATA
Text Categorization
Information
Retrieval
Data Mining
Genome Sequence
Matching
What is a pattern?
“A pattern is the opposite of a chaos; it is an entity
vaguely defined, that could be given a name.”
PR Definitions
Theory, Algorithms, Systems to Put
Patterns into Categories
Classification of Noisy or Complex Data
Relate Perceived Pattern to Previously
Perceived Patterns
Characters
A v t u I h D U w K
Ç ş ğ İ üÜ Ö Ğ
‫چك‬٤٧‫ع‬
К Ц Д
ζω Ψ Ω ξ θ
‫נ‬
‫ד‬
‫ת‬
‫ש‬
‫א ם‬
Handwriting
Terminology
Features, feature vector
Decision boundary
Error
Cost of error
Generalization
A Fishy Example I
“Sorting incoming Fish on a conveyor
according to species using optical sensing”
Salmon or Sea Bass?
Problem Analysis

Set up a camera and take some sample images to extract
features





Length
Lightness
Width
Number and shape of fins
Position of the mouth, etc…
This is the set of all suggested features to explore for use in our
classifier!
Solution by Stages
Preprocess raw data from camera
Segment isolated fish
Extract features from each fish (length,width,
brightness, etc.)
Classify each fish
Preprocessing
Use a segmentation operation to isolate fishes
from one another and from the background
Information from a single fish is sent to a feature
extractor whose purpose is to reduce the data by
measuring certain features

The features are passed to a classifier
2
2
Classification
Select the length of the fish as a possible
feature for discrimination
2
2
The length is a poor feature alone!
Select the lightness as a possible feature.
2
2
“Customers do not want sea
bass in their cans of salmon”
Threshold decision boundary and cost relationship
 Move our decision boundary toward smaller values
of lightness in order to minimize the cost (reduce
the number of sea bass that are classified salmon!)

Task of decision theory
2
Adopt the lightness and add the width of
the fish
Fish
x = [x1, x2]
Width
Lightness
2
2
We might add other features that are not
correlated with the ones we already have. A
precaution should be taken not to reduce the
performance by adding such “noisy features”
Ideally, the best decision boundary should be
the one which provides an optimal performance
such as in the following figure:
2
2
However, our satisfaction is
premature because the central
aim of designing a classifier is to
correctly classify novel input
Issue of generalization!
2
2
Decision Boundaries
Observe: Can do much better with two features
Caveat: overfitting!
Occam’s Razor
Entities are not to be multiplied without necessity
William of Occam
(1284-1347)
A Complete PR System
Problem Formulation
Input
object
Measurements
&
Preprocessing
Class
Label
Features
Classification
Basic ingredients:
•Measurement space (e.g., image intensity, pressure)
•Features (e.g., corners, spectral energy)
•Classifier - soft and hard
•Decision boundary
•Training sample
•Probability of error
Pattern Recognition
Systems
Sensing
Use of a transducer (camera or
microphone)
 PR system depends of the bandwidth, the
resolution, sensitivity, distortion of the
transducer

Segmentation and grouping

Patterns should be well separated and
should not overlap
3
3
Feature extraction


Discriminative features
Invariant features with respect to translation, rotation and
scale.
Classification

Use a feature vector provided by a feature extractor to
assign the object to a category
Post Processing

Exploit context dependent information other than from the
target pattern itself to improve performance
The Design Cycle
Data collection
Feature Choice
Model Choice
Training
Evaluation
Computational Complexity
4
4
Data Collection
How do we know when we have collected an
adequately large and representative set of
examples for training and testing the system?
4
Feature Choice
Depends on the characteristics of the
problem domain. Simple to extract,
invariant to irrelevant transformation
insensitive to noise.
4
Model Choice
Unsatisfied with the performance of our linear fish
classifier and want to jump to another class of
model
4
Training
Use data to determine the classifier. Many
different procedures for training classifiers and
choosing models
4
Evaluation
Measure the error rate (or performance) and
switch from one set of features & models to
another one.
4
Computational Complexity
What is the trade off between computational ease
and performance?
(How an algorithm scales as a function of the
number of features, number or training examples,
number patterns or categories?)
4
Learning and Adaptation
Learning: Any method that combines empirical information from
the environment with prior knowledge into the design of a
classifier, attempting to improve performance with time.
Empirical information: Usually in the form of training examples.
Prior knowledge: Invariances, correlations
Supervised learning

A teacher provides a category label or cost for each pattern in the
training set
Unsupervised learning

The system forms clusters or “natural groupings” of the input patterns
5
Syntactic Versus Statistical
PR
Basic assumption: There is an underlying regularity
behind the observed phenomena.
Question: Based on noisy observations, what is the
underlying regularity?
Syntactic: Structure through common generative
mechanism. For example, all different manifestations
of English, share a common underlying set of
grammatical rules.
Statistical: Objects characterized through statistical
similarity. For example, all possible digits `2' share
some common underlying statistical relationship.
Difficulties
Segmentation
Context
Temporal structure
Missing features
Aberrant data
Noise
Do all these images represent an `A'?
Design Cycle
How do we know what features to select, and how do we select
them…?
What type of classifier shall we use. Is there best classifier…?
How do we train…?
How do we combine prior knowledge with
empirical data?
How do we evaluate our performance
Validate the results. Confidence in decision?
Conclusion
I expect you are overwhelmed by the number,
complexity and magnitude of the subproblems of Pattern Recognition
Many of these sub-problems can indeed be
solved
Many fascinating unsolved problems still
remain
6
Toolkit for PR
Statistics
Decision Theory
Optimization
Signal Processing
Neural Networks
Fuzzy Logic
Decision Trees
Clustering
Genetic Algorithms
AI Search
Formal Grammars
….
Linear algebra
Matrix A:
A  [aij ]mn
 a11 a12
a
a22
21


 ...
...

am1 am 2
... a1n 
... a2 n 
... ... 

... amn 
Matrix Transpose
B  [bij ]nm  AT  bij  a ji ; 1  i  n, 1  j  m
Vector a
 a1 
a   ... ; aT  [a1 ,..., an ]
an 
Matrix and vector multiplication
Matrix multiplication
A  [aij ]m p ; B  [bij ] pn ;
AB  C  [cij ]mn ,where cij  rowi ( A)  col j ( B)
Outer vector product
a  A  [aij ]m1; bT  B  [bij ]1n ;
c  a  b  AB, an m  n matrix
Vector-matrix product
A  [aij ]mn ; b  B  [bij ]n1 ;
C  Ab  an m 1 matrix  vector of length m
Inner Product
n
a T  b   ai bi
Inner (dot) product:
i 1
Length (Eucledian norm) of a vector
a is normalized iff ||a|| = 1
a  aT  a 
n
 ai
i 1
aT  b
cos  
|| a || || b ||
The angle between two ndimesional vectors
An inner product is a measure of
collinearity:
T

a and b are orthogonal iff

a and b are collinear iff
a b  0
a  b || a || || b ||
A set of vectors is linearly
independent if no vector is a linear
combination of other vectors.
T
2
Determinant and Trace
A  [aij ]nn ;
n
Determinant
det( A)   aij Aij ; i  1,....n;
j 1
Aij  (1)i  j det( M ij )
det(AB)= det(A)det(B)
Trace
n
A  [aij ]nn ; tr[ A]   a jj
j 1
Matrix Inversion
A (n x n) is nonsingular if there
exists B
AB  BA  I n ; B  A1
A=[2 3; 2 2], B=[-1 3/2; 1 -1]
|| A || 0
A is nonsingular iff
Pseudo-inverse for a non square
matrix, provided
#
T
1 T
A

[
A
A
]
A
T
A A is not singular
A A I
#
Eigenvectors and Eigenvalues
Ae j   j e j , j  1,..., n; || e j || 1
Characteristic equation:
n-th order polynomial, with n roots.
n
tr[ A]    j
j 1
det[ A  I n ]  0
n
det[ A]    j
j 1
Probability Theory
Primary references:


Any Probability and Statistics text book (Papoulis)
Appendix A.4 in “Pattern Classification” by Duda
et al
The principles of probability theory, describing
the behavior of systems with random
characteristics, are of fundamental
importance to pattern recognition.
Example 1 ( wikipedia)
•two bowls full of cookies.
•Bowl #1 has 10 chocolate chip cookies and 30 plain
cookies,
•bowl #2 has 20 of each.
•Fred picks a bowl at random, and then picks a cookie at
random.
•The cookie turns out to be a plain one.
•How probable is it that Fred picked it out of bowl
•what’s the probability that Fred picked bowl #1, given
that he has a plain cookie?”
•event A is that Fred picked bowl #1,
•event B is that Fred picked a plain cookie.
•Pr(A|B) ?
Example1 - cpntinued
Tables of occurrences and relative frequencies
It is often helpful when calculating conditional probabilities to create a simple table containing the number of occurrences of each outcome, or
the relative frequencies of each outcome, for each of the independent variables. The tables below illustrate the use of this method for the
cookies.
Relative frequency of cookies in each bowl
by type of cookie
Number of cookies in each bowl
by type of cookie
Bowl 1
Bowl 2
Totals
Bowl
#1
Bowl
#2
Totals
0.125
0.250
0.375
Chocolate Chip
10
20
30
Chocolate
Chip
Plain
30
20
50
Plain
0.375
0.250
0.625
Total
40
40
80
Total
0.500
0.500
1.000
The table on the right is derived from the table on the left by dividing each entry by the total number of cookies under consideration, or 80 cookies.
Example 2
X
Y
Z
P(x,y,z)
0
0
0
0
1
1
1
1
0
0
1
1
0
0
1
1
0
1
0
1
0
1
0
1
0.07
0.04
0.03
0.18
0.16
0.18
0.21
0.13
1. Power Plant Operation.



The variables X, Y, Z describe
the state of 3 power plants
(X=0 means plant X is idle).
Denote by A an event that a
plant X is idle, and by B an
event that 2 out of three plants
are working.
What’s P(A) and P(A|B), the
probability that X is idle given
that at least 2 out of three are
working?
P(A) = P(0,0,0) + P(0,0,1) + P(0,1,0) +
P(0, 1, 1) = 0.07+0.04 +0.03 +0.18
=0.32
P(B) = P(0,1,1) +P(1,0,1) + P(1,1,0)+
P(1,1,1)= 0.18+ 0.18+0.21+0.13=0.7
P(A and B) = P(0,1,1) = 0.18
P(A|B) = P(A and B)/P(B) = 0.18/0.7
=0.257
2. Cars are assembled in four possible
locations. Plant I supplies 20% of the cars;
plant II, 24%; plant III, 25%; and plant IV,
31%. There is 1 year warrantee on every car.
The company collected data that shows
P(claim| plant I) = 0.05; P(claim|Plant II)=0.11;
P(claim|plant III) = 0.03; P(claim|Plant IV)=0.18;
Cars are sold at random.
An owned just submitted a claim for her car.
What are the posterior probabilities that this
car was made in plant I, II, III and IV?
P(claim) = P(claim|plant I)P(plant I) +
P(claim|plant II)P(plant II) +
P(claim|plant III)P(plant III) +
P(claim|plant IV)P(plant IV) =0.0687
P(plant1|claim) =
= P(claim|plant I) * P(plant I)/P(claim) = 0.146
P(plantII|claim) =
= P(claim|plant II) * P(plant II)/P(claim) = 0.384
P(plantIII|claim) =
= P(claim|plant III) * P(plant III)/P(claim) = 0.109
P(plantIV|claim) =
= P(claim|plant IV) * P(plant IV)/P(claim) = 0.361
Example 3
3. It is known that 1% of population suffers from a
particular disease. A blood test has a 97% chance to
identify the disease for a diseased individual, by also
has a 6% chance of falsely indicating that a healthy
person has a disease.
a. What is the probability that a random person has a
positive blood test.
b. If a blood test is positive, what’s the probability that
the person has the disease?
c. If a blood test is negative, what’s the probability that
the person does not have the disease?
A is the event that a person has a disease. P(A) =
0.01; P(A’) = 0.99.
B is the event that the test result is positive.


P(B|A) = 0.97; P(B’|A) = 0.03;
P(B|A’) = 0.06; P(B’|A’) = 0.94;
(a) P(B) = P(A) P(B|A) + P(A’)P(B|A’) = 0.01*0.97
+0.99 * 0.06 = 0.0691
(b) P(A|B)=P(B|A)*P(A)/P(B) = 0.97* 0.01/0.0691 =
0.1403
(c) P(A’|B’) = P(B’|A’)P(A’)/P(B’)= P(B’|A’)P(A’)/(1P(B))= 0.94*0.99/(1-.0691)=0.9997
Sums of Random Variables
z=x+y
z  x   y
Var(z) = Var(x) + Var(y) + 2Cov(x,y)
If x,y independent: Var(z) = Var(x) + Var(y)
Distribution of z:

p( z )  p x ( x)  p y ( y ) 
 p ( x) p
x

y
( z  x)dx
Examples:
x and y are uniform on [0,1]

Find p(z=x+y), E(z), Var(z);
x is uniform on [-1,1], and P(y)= 0.5 for
y =0, y=10; and 0 elsewhere.

Find p(z=x+y), E(z), Var(z);
Normal Distributions
Gaussian distribution
p( x)  N ( , ) 
x
Mean
Variance
x
1
2  x
e
( x   x ) 2 / 2 x 2
E ( x)   x
E[( x  x )2 ]   x2
Central Limit Theorem says sums of random variables tend
toward a Normal distribution.
Mahalanobis Distance:
r
x x
x
Multivariate Normal Density
x is a vector of d Gaussian variables
p( x)  N ( ,) 
1
2 d / 2 ||1 / 2
1
 ( x   )T  1( x   )
e 2

  E[ x]  xp( x)dx


  E[( x   )( x   ) ]  ( x   )( x   )T p( x)dx

T
Mahalanobis Distance
r 2  ( x  )T 1( x  )
All conditionals and marginals are also Gaussian
Bivariate Normal Densities
Level curves - elliplses.
x and y width are determined by the
variances, and the eccentricity by
correlation coefficient
 Principal axes are the eigenvectors, and
the width in these direction is the root of
the corresponding eigenvalue.

Information theory
Key principles:

What is the information contained in a
random event?


Less probable event contains more information
For two independent event, the information is a sum
I ( x)   log 2 P( x)

What is the average information or entropy
of a distribution?
H ( x)   P( x) log 2 P( x)
x

Examples: uniform distribution, dirac
distribution;

Mutual information: reduction in uncertainty
about one variable due to knowledge of other
variable.
p ( x, y )
Ix , y  H ( x)  H ( x | y )   p( x, y ) log 2
p( x) p( y )
x, y
Download