1. Introduction to Machine
Learning
Lecture 3
Albert Orriols i Puig
aorriols@salle.url.edu
i l @ ll ld
Artificial Intelligence – Machine Learning
Enginyeria i Arquitectura La Salle
gy q
Universitat Ramon Llull
2. Recap of Lecture 2
Machine learning
Learning = Improving with experience at some task
Improve over task T
With respect to a performance measure P
Based on experience E
Three especial niches
Data mining: extract information from historical data to help
g p
decision making
Software applications that are too complex to build a hard-
pp p
wired solution for
Self customizing p g
g programs
Slide 2
Artificial Intelligence Machine Learning
3. Today’s Agenda
Characteristics Desired for ML Methods
General issues
Concepts that will be used through lectures
Summary of the Paradigms that We Won’t
y g
Study
Summary of the P bl
S f th Problems th t W Will Study
that We St d
Slide 3
Artificial Intelligence Machine Learning
4. Characteristics Desired ML
We would like our ML techniques to have the following
q g
properties
Be able to generalize but not too much
generalize,
Be robust
Be li bl
B reliable
Learn models of high quality
Be scalable and efficient
Be explicative
Be determinist
Slide 4
Artificial Intelligence Machine Learning
5. Characteristics Desired ML
Be able to generalize, but not too much
g ,
We learn from a set of examples
Imagine that we are doing d t regression
I i th t d i data i
Examples (observations)
-- Real domain
Learned function
We only know the examples {e1, e2, e3, e4, e5, e6, e7, e8, e9}
We do not know the real distribution
So, does the learning function fits t e real d st but o
t e ea g u ct o ts the ea distribution?
Slide 5
Artificial Intelligence Machine Learning
6. Characteristics Desired ML
Be able to generalize, but not too much
g ,
Examples (observations)
-- Real domain
Learned function
What could have happened?
at cou d a e appe ed
I may not be a good representation of the original distribution
The ML method may not work well (overfitting)
So, what should we do?
Assume that I is a good representative of the original distribution
g p g
Go for the simplest solution
Slide 6
Artificial Intelligence Machine Learning
7. Characteristics Desired ML
Be robust
Real-world is imperfect and our measurements of real world
may be e e more imperfect
ay even o e pe ec
Therefore, we will deal with domains with
Noise
Uncertainty
Vagueness
We have to keep this in mind when designing our algorithms
Slide 7
Artificial Intelligence Machine Learning
8. Characteristics Desired ML
Learn models of high quality
gq y Test set
How do we evaluate learning quality? New instance
Information based Knowledge
on experience extraction
Learner Model
Dataset
Predicted Output
Training set
g
More advanced validation methods:
k-fold cross-validation
Holdout
Slide 8
Artificial Intelligence Machine Learning
9. Characteristics Desired ML
Be reliable
What do you prefer?
Do not predict something that you doubt about?
Or just bet for an option?
Classes are cost sensitive?
Cl t iti ?
What happens if I say that a patient, who has actually cancer,
is healthy?
What happens if I say that a patient, who is actually healthy,
has cancer?
Do I prefer to model one class as opposed to the other?
Fraud detection (0.1% of fraudulent transactions)
(% )
Geez, I modeled perfectly the non-fraudulent transactions!
Am I successful?
Slide 9
Artificial Intelligence Machine Learning
10. Characteristics Desired ML
Be scalable and efficient
Huge amount of data
Information hidden i th
If ti hidd in these d t
data
I need to process them quickly!
Two types o costs
o of costs:
Cost to build the model
Cost to classify new test examples
y p
Slide 10
Artificial Intelligence Machine Learning
11. Characteristics Desired ML
Be explicative
Should
Sh ld I care about giving an explanation?
b t ii l ti ?
Text/speech recognition
fast. huge,
Things happen too fast If errors are not too huge I do not care if
I read “a” instead of “e”
Medical diagnosis
g
I really care about obtaining an accurate explanation, since the
diagnosis may involve applying surgery to a patient or not
Slide 11
Artificial Intelligence Machine Learning
12. Characteristics Desired ML
Be determinist
If my data does not change
The learned model should be always the same
The answer for a given test instance should be always the
same
If my data changes
I should adapt to the changes
Slide 12
Artificial Intelligence Machine Learning
13. Paradigms in ML
Typically, techniques in ML have been divided in
different paradigms
Inductive learning
Explanation-based learning
p g
Analogy-based learning
Evolutionary learning
Connectionist Learning
Slide 13
Artificial Intelligence Machine Learning
14. Inductive Learning
Induce rules, trees or, in general, patterns from a set of
, ,g ,p
examples
Start from a specific experience
Draw inferences or generalizations from it
That is
Initial state: Original data
State: Symbolic description of the data with a certain degree of
generalization/specialization
Final state: Model with maximum generalization that implies the
input data
Slide 14
Artificial Intelligence Machine Learning
15. Explanation-Based Learning
Deduce information from a set of observations
Humans learn a lot from few examples
Machine: use results f
M hi lt from one example t solve th next
l to l the t
problem
Domain theory for the problem
EBL New domain theory
Goal concept
Training example
Slide 15
Artificial Intelligence Machine Learning
17. Explanation-based Learning
Example
Goal: Get to Brecon
Training data
Near (Cardiff, Brecon)
Airport (Cardiff)
Domain Knowledge
Near(x,y) ^ holds( loc(x), s ) holds( loc(y), result(drive(x,y),s) )
Airport(z) loc(z),
loc(z) result( fly(z), s )
fly(z)
Operational criterion: We must express concept definition in pure description
language syntax
Our goal can be expressed as
Holds ( loc(Brecon), s)
Slide 17
Artificial Intelligence Machine Learning
18. Learning Based on Analogy
A is similar to A’ according to α
α
A A’
If I have B, can I get B’?
Learn the causality relationship β β
β'
β
Transform α to α’
α'
B B’
Get B according to B and α’
B’ α
Where is the trick?
In learning α’ and β
Partial mapping
Previously
New Problem
solved problem
Derivation
Transformation
Solution to the Solution of this
problem known problem
Slide 18
Artificial Intelligence Machine Learning
19. Evolutionary Learning
Nature as problem solver
p
Nature evolved adapted solutions to life
Let’s
L t’ use thi concepts t learn f
this t to l from experience
i
Slide 19
Artificial Intelligence Machine Learning
20. Connectionist Learning
Mimic brain structure to build machines that are able to
learn
A brain consists of
Connected neurons that behave in a specific way
Let’s assume that this behavior can be coded functionally
Slide 20
Artificial Intelligence Machine Learning
21. Problems That We’ll Study
Typical ML courses go through the different families
yp g g
Structured courses
Big i t
Bi picture of th diff
f the different l
t learning paradigms
i di
However
Emergence of hybrid intelligent systems
Concepts come all mixed together
g
We are engineers. We need to solve problems
So,
So we propose to go problem-oriented
problem oriented
Techniques of different paradigms will come on our way
Slide 21
Artificial Intelligence Machine Learning
22. Problems That We’ll Study
Data classification: C4.5, kNN, Naïve Bayes …
1.
Statistical learning: SVM
2.
2
Association analysis: A-priori
3.
Link mining: Page Rank
4.
Clustering: k-means
g
5.
Reinforcement learning: Q-learning, XCS
6.
Regression
7.
7
Genetic Fuzzy Systems
8.
Slide 22
Artificial Intelligence Machine Learning
23. Next Class
How I Would Like my Problem to Look Like?
Summary of the Paradigms that we Won’t Study
Slide 23
Artificial Intelligence Machine Learning
24. Introduction to Machine
Learning
Lecture 3
Albert Orriols i Puig
aorriols@salle.url.edu
i l @ ll ld
Artificial Intelligence – Machine Learning
Enginyeria i Arquitectura La Salle
gy q
Universitat Ramon Llull