lOMoARcPSD|17897360
Lecture notes, lecture 2 - Basic Linear Algebra - Matlab
Artificial Inteligence II (The University of Western Ontario)
StuDocu is not sponsored or endorsed by any college or university
Downloaded by Marisnelvys Cabreja (
[email protected])
lOMoARcPSD|17897360
CS4442/9542b
Artificial Intelligence II
Prof. Olga Veksler
Lecture 2
Introduction to ML
Basic Linear Algebra
Matlab
Some slides on Linear Algebra are from Patrick Nichols
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Outline
• Introduction to Machine Learning
• Basic Linear Algebra
• Matlab Intro
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Intro: What is Machine Learning?
• How to write a computer program that automatically
improves its performance through experience
• Machine learning is useful when it is too difficult to
come up with a program to perform a desired task
• Make computer to learn by showing examples (most
frequently with correct answers)
• “supervised” learning or learning with a teacher
• In practice: computer program (or function) which has
a tunable parameters, tune parameters until the
desirable behavior on the examples
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Different Types of Learning
• Learning from examples:
• Supervised Learning: given training examples of
inputs and corresponding outputs, produce the
“correct” outputs for new inputs
• study in this course
• Unsupervised Learning: given only inputs as
training, find structure in the world: e.g. discover
clusters
• Other types, such as reinforcement learning are
not covered in this course
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Supervised Machine Learning
• Training samples (or examples) x1,x2,…, xn
• Each example xi is typically multi-dimensional
• xi1, xi2 ,…, xid are called features, xi is often called a
feature vector
• Example: x1 = {3,7, 35}, x2 = {5, 9, 47}, …
• how many and which features do we take?
• Know desired output for each example y1, y2,…yn
• This learning is supervised (“teacher” gives desired outputs)
• yi are often one-dimensional
• Example: y1 = 1 (“face”), y2 = 0 (“not a face”)
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Supervised Machine Learning
• Two types of supervised learning:
• Classification (we will only do classification in this
course):
• yi takes value in finite set, typically called a label
or a class
• Example: yi ∈{“sunny”, ”cloudy”, ”raining”}
• Regression
• yi continuous, typically called an output value
• Example: yi = temperature ∈[-60,60]
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Toy Application: fish sorting
classifier
fis
h sp
fis ec
h im ie s salmon
ag
e
sorting
chamber
sea bass
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Classifier design
• Notice salmon tends to be shorter than sea bass
• Use fish length as the discriminating feature
• Count number of bass and salmon of each length
2 4 8 10 12 14
bass 0 1 3 8 10 5
salmon 2 5 10 5 1 0
12
10
8
Count
salmon
6
sea bass
4
2
0
2 4 8 10 12 14
Length
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Single Feature (length) Classifier
• Find the best length L threshold
fish length < L fish length > L
classify as salmon classify as sea bass
• For example, at L = 5, misclassified:
• 1 sea bass 2 4 8 10 12 14
bass 0 1 3 8 10 5
• 16 salmon
salmon 2 5 10 5 1 0
• Classification error (total error) 17 = 34%
50
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Single Feature (length) Classifier
fish classified fish classified
as salmon as sea bass
12
10
8
Count
salmon
6
sea bass
4
2
0
2 4 8 10 12 14
Length
• After searching through all possible thresholds L, the
best L= 9, and still 20% of fish is misclassified
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Next Step
• Lesson learned:
• Length is a poor feature alone!
• What to do?
• Try another feature
• Salmon tends to be lighter
• Try average fish lightness
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Single Feature (lightness) Classifier
1 2 3 4 5
bass 0 1 2 10 12
salmon 6 10 6 1 0
14
12
10
Count
8 salmon
6 sea bass
4
2
0
1 2 3 4 5
Lightness
• Now fish are classified best at lightness
threshold of 3.5 with classification error of 8%
Downloaded by Marisnelvys Cabreja (
[email protected])
lOMoARcPSD|17897360
Can do better by feature combining
• Use both length and lightness features
• Feature vector [length,lightness]
ba decision
ss
lightness
boundary
decision regions
sa
lm
on
length
• Classification error 4%
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Even Better Decision Boundary
lightness
length
• Decision boundary (wiggly) with 0% classification error
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Test Classifier on New Data
• The goal is for classifier to perform well on new data
• Test “wiggly” classifier on new data: 25% error
lightness
length
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
What Went Wrong?
added 2 samples
• We always have only a limited amount of data, not all
possible data
• We should make sure the decision boundary does not
adapt too closely to the particulars of the data we have
at hand, but rather grasps the “big picture”
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
What Went Wrong: Overfitting
• Complicated boundaries overfit the data, they are too
tuned to the particular training data at hand
• Therefore complicated boundaries tend to not
generalize well to the new data
• We usually refer to the new data as “test” data
Downloaded by Marisnelvys Cabreja (
[email protected])
lOMoARcPSD|17897360
Overfitting: Extreme Example
• Say we have 2 classes: face and non-face images
• Memorize (i.e. store) all the “face” images
• For a new image, see if it is one of the stored faces
• if yes, output “face” as the classification result
• If no, output “non-face”
• also called “rote learning”
• problem: new “face” images are different from stored
“face” examples
• zero error on stored data, 50% error on test (new) data
• Rote learning is memorization without generalization
slide is modified from Y. LeCun
Downloaded by Marisnelvys Cabreja (
[email protected])
lOMoARcPSD|17897360
Generalization
training data test data
• The ability to produce correct outputs on previously unseen
examples is called generalization
• The big question of learning theory: how to get good generalization
with a limited number of examples
• Intuitive idea: favor simpler classifiers
• William of Occam (1284-1347): “entities are not to be multiplied without necessity”
• Simpler decision boundary may not fit ideally to the training data
but tends to generalize better to new data
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Underfitting
• We can also underfit data, i.e. use too simple decision
boundary
• chosen model is not expressive enough
• There is no way to fit a linear decision boundary so that
the training examples are well separated
• Training error is too high
• test error is, of course, also high
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Underfitting Overfitting
underfitting “just right” overfitting
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Sketch of Supervised Machine Learning
• Chose a learning machine f(x,w)
• w are tunable weights
• x is the input sample
• f(x,w) should output the correct class of sample x
• use labeled samples to tune weights w so that f(x,w)
give the correct label for sample x
• Which function f(x,w) do we choose?
• has to be expressive enough to model our problem
well, i.e. to avoid underfitting
• yet not to complicated to avoid overfitting
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Training and Testing
• There are 2 phases, training and testing
• Divide all labeled samples x1,x2,…xn into 2 sets,
training set and test set
• Training phase is for “teaching” our machine (finding
optimal weights w)
• Testing phase is for evaluating how well our machine
works on unseen examples
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Training Phase
• Find the weights w s.t. f(xi,w) = yi “as much as
possible” for training samples (xi, yi)
• “as much as possible” needs to be defined
• How do we find parameters w to ensure
f(xi,w) = yi for most training samples (xi,yi) ?
• This step is usually done by optimization, can be
quite time consuming
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Testing Phase
• The goal is to design machine which performs
well on unseen examples
• Evaluate the performance of the trained
machine f(x,w) on the test samples (unseen
labeled samples)
• Testing the machine on unseen labeled examples
lets us approximate how well it will perform in
practice
• If testing results are poor, may have to go back
to the training phase and redesign f(x,w)
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Generalization and Overfitting
• Generalization is the ability to produce correct
output on previously unseen examples
• In other words, low error on unseen examples
• Good generalization is the main goal of ML
• Low training error does not necessarily imply that
we will have low test error
• we have seen that it is easy to produce f(x,w) which is
perfect on training samples (rote “learning”)
• Overfitting
• when the machine performs well on training data but
poorly on test data
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Classification System Design Overview
• Collect and label data by hand
salmon sea bass salmon salmon sea bass sea bass
• Split data into training and test sets
• Preprocess by segmenting fish from background
• Extract possibly discriminating features
• length, lightness, width, number of fins,etc.
• Classifier design
• Choose model for classifier we look at these two
• Train classifier on training data steps in this course
• Test classifier on test data
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Basic Linear Algebra
• Basic Concepts in Linear Algebra
• vectors and matrices
• products and norms
• vector spaces and linear transformations
• Introduction to Matlab
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Why Linear Algebra?
• For each example (e.g. a fish image), we extract a set
of features (e.g. length, width, color)
• This set of features is represented as a feature vector
• [length, width, color]
• All collected examples will be represented as
collection of (feature) vectors
[l1, w1 , c1 ] example 1 l1 w1 c1
[l2 , w2 , c2 ] example 2 l2 w2 c2
[l3 , w3 , c3 ] example 3
l3 w3 c3
matrix
• Also, we will use linear models since they are simple
and computationally tractable
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
What is a Matrix?
• A matrix is a set of elements, organized into
rows and columns
rows
2 7 6 10 example 1
columns 1 4 4 9 example 2
6 4 9 feature 2 6 example 3
feature 3
feature 4
feature 1
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Basic Matrix Operations
• addition, subtraction, multiplication by a scalar
a b e f a+e b+ f
+ = add elements
c d g h c+g d +h
a b e f a−e b− f
− = subtract elements
c d g h c−g d −h
α ⋅ a α ⋅b
a b
α⋅ = multiply every entry
c d α ⋅c α ⋅d
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Matrix Transpose
T
• n by m matrix A and its m by n transpose A
x11 x12 x1m x11 x21 xn1
x21 x22 x2m T
x12 x22 xn2
A= A =
xn1 xn2 xnm x1m x2m xnm
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Vectors
• Vector: N x 1 matrix
x1
v=
x2
• dot product and magnitude defined on vectors only
x2 x2 x2
v a a+b a a-b
b b
x1 x1 x1
vector addition vector subtraction
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
More on Vectors
• n-dimensional row vector x = [x1 x2 xn ]
x1
T
x2
• Transpose of row vector is column vector x =
xn
• Vector product (or inner or dot product)
x, y = x ⋅ y = xT y = x1 y1 + x2 y2 + + xn yn = xi yi
i=1 n
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
More on Vectors
• Euclidian norm or length x = x, x = xi2
i =1 n
• If x =1 we say x is normalized or unit length
xT y
• angle q between vectors x and y : cos θ =
x y
• inner product captures direction relationship
cos θ = 0 cos θ = 1 cos θ = −1
y y y
x
x
T
x
x y=0 xT y = x y > 0 xT y = − x y < 0
x⊥y
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
More on Vectors
• Vectors x and y are orthonormal if they are
orthogonal and x = y =1
• Euclidian distance between vectors x and y
x x-y
x− y = (xi − yi ) 2
i =1 n
y
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Linear Dependence and Independence
• Vectors x1, x2,…, xn are linearly dependent if
there exist constants α1, α2,…, αn s.t.
• α1x1+ α2x2+…+αnxn = 0
• αi 0 for at least one I
• Vectors x1, x2,…, xn are linearly independent if
α1x1+ α2x2+…+αnxn = 0 α1 = α2=…= αn= 0
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Vector Spaces and Basis
• The set of all n-dimensional vectors is called a
vector space V
• A set of vectors {u1,u2,…, un } are called a basis
for vector space if any v in V can be written as
v = α1u1+ α2u2+…+αnun
• u1,u2,…, un are independent implies they form a
basis, and vice versa
• u1,u2,…, un give an orthonormal basis if
1. ui = 1 ∀i
2. ui ⊥ u j ∀i ≠ j
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Orthonormal Basis
• x, y,…, z form an orthonormal basis
x = [1 0 0] x⋅ y = 0
T
y = [0 1 0]
T x⋅z = 0
z = [0 0 1]
T y⋅z = 0
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Matrix Product
b11 b1m
a11 a12 a13 a1d b21 b2m
AB = b31 b3m = cij
an1 an2 an3 and
bd1 bdm cij = ai, bj
ai is row i of A
bj is column j of B
• # of columns of A = # of rows of B
• even if defined, in general AB BA
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Matrices
• Rank of a matrix is the number of linearly
independent rows (or equivalently columns)
• A square matrix is non-singular if its rank equal
to the number of rows. If its rank is less than
number of rows it is singular.
1 0 0
0 1 0
• Identity matrix I=
0 0
AI=IA=A 1 2 9 5
0 0 1
T
2 7 4 8
• Matrix A is symmetric if A=A
9 4 3 6
5 8 6 4
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
Matrices
-1
• Inverse of a square matrix A is matrix A s.t.
-1
AA = I
• If A is singular or not square, inverse does not
exist
T
• Pseudo-inverse A is defined whenever A A is
not singular (it is square)
T -1 T
A = (A A) A
T -1 T
AA =(A A) AA=I
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
MATLAB
Downloaded by Marisnelvys Cabreja ([email protected])
lOMoARcPSD|17897360
• Starting matlab • Elementary functions
• xterm -fn 12X24 • help elfun
• matlab • Data types
• Basic Navigation • double
• quit • Char
• more • Programming in Matlab
• help general • .m files
• Scalars, variables, basic arithmetic • scripts
• Clear • function y=square(x)
• + - */ ^ • help lang
• help arith • Flow control
• Relational operators • if i== 1else end, if else if end
• ==,&,|,~,xor • for i=1:0.5:2 … end
• help relop • while i == 1 … end
• Lists, vectors, matrices • Return
• A=[2 3;4 5] • help lang
• A’ • Graphics
• Matrix and vector operations • help graphics
• find(A>3), colon operator • help graph3d
• * / ^ .* ./ .^ • File I/O
• eye(n),norm(A),det(A),eig(A) • load,save
• max,min,std • fopen, fclose, fprintf, fscanf
• help matfun
Downloaded by Marisnelvys Cabreja ([email protected])
