Artificial Intelligence

4. Machine Learning

4.1. Introduction to machine learning

Dr. Xavi Varona, Sept. 2019

Artificial Intelligence?
1. a branch of computer science dealing with the simulation of intelligent
behaviour in computers
2. the capability of a machine to imitate intelligent human behaviour
AI technologies
It is a set of rich sub-disciplines
and methods (machine
learning, pattern recognition,
computer vision,…).

We have to consider all these

different disciplines and
methods in seeking true
solutions in delivering value to
human beings and
organizations

Eric Horvitz (Microsoft Research)

AI & Machine Learning
ML development
Symbolic AI vs Machine Learning
Methods in artificial
intelligence research that are
based on high-level "symbolic"
(human-readable)
representations of
problems, logic and search.

Symbolic AI was the

dominant paradigm of AI
research from the mid-1950s (blueberry (is a fruit)
until the late 1980s (shape round)
(color purple)
(size .4 inch))
Machine Learning
Machine learning is a subset of artificial intelligence in the field of computer
science that often uses statistical techniques to give computers the ability
to "learn" (i.e., progressively improve performance on a specific task)
with data, without being explicitly programmed.
Machine Learning?
Without learning, the agents
must be AI programmed during
the design. Therefore the
machine intelligence depends
from the designer and his ability
to predict the diversity of the
environment and its possible
changes.

Is it possible?
ML & Deep Learning
Neural Networks / Deep learning
Why Machine Learning in CS?

11
Big data, the big bang
Expansion
ML “definition”
A computer program is said
to learn from experience E with
respect to some class of
tasks T and performance
measure P, if its performance at
tasks in T, as measured by P,
improves with experience E.

Tom Mitchell (CMU, 1998)

Task: defining the problem

Example: Gender recognition

male female
Experience: Data
 Learning (training)

updating
model…
female
male
male
female
female
male
Testing (probability)

This is a female
( 98% sure)
?
Testing (probability)

This is a male
( 70% sure)
?
Task types
Classification

Classification algorithms are

used when the desired output is
a discrete label.

Regression

Approximate a continuous
function (predicting outputs
that are continuous)
ML “definition”
A computer program is said
to learn from experience E with
respect to some class of
tasks T and performance
measure P, if its performance at
tasks in T, as measured by P,
improves with experience E.

Tom Mitchell (CMU, 1998)

 Ex: Task
EX: Suposem que el vostre programa de correu electrònic veu els missatges que heu o no
heu marcat com a correu brossa, i a partir d’això aprèn com millorar el seu filtre d’spam.

Quina és la T en aquest context?

q Veure els missatges que l’usuari etiqueta

q Classificar els missatges com a correu brossa o desitjat.

q El nombre (o fracció) de missatges de correu electrònic classificats correctament com a

correu brossa / correu desitjat.

q Cap de les anteriors, això no és un problema d'aprenentatge computacional.

Experience
The experience, E is the data to
be collect in order to learn the
task.

Features: language in which we

describe the properties of
objects in a domain; You have
to choose them, pre-process
them, manipulate them ... (data
mining)
ML problem statement: train

inputs predictions
Learning algorithm
Supervised (classification,
regression): they are provided
(through a "master") pairs of
cases-labels, indicating whether
or not they belong to the type of
association to be learned:
prediction

Unsupervised (clustering): the

agent must use other funds to
obtain a "feedback" that tells him
whether he does it well or not,
since he does not have a
"master": description
 Ex: Learning algorithm

EX: En els següents exemples, quan aplicaríeu aprenentatge no supervisat?

q Donat correus etiquetats com a spam/no spam, aprendre un filtre d’spam

q Donades un conjunt d’articles trobats a la web, agrupar-los en conjunts que conten la

mateixa història.

q Donada una base de dades d’informació sobre consumidors, agrupar-los

automàticament en segments diferents de mercat.

q Donada una base de dades de pacients diagnosticats amb diabetis o no, aprendre a
classificar nous pacients segons si tenen diabetis o no.
Performance (Learning)

• Training dataset targets

• Supplied test set

• Percentage split

• Cross validation

inputs predictions
ML problem statement: train

inputs predictions
ML problem statement: predict

inputs predictions
Performance
Classification Task:

• True positive (TP) eqv. with hit

• True negative (TN) eqv. with

correct rejection

• False positive (FP) eqv. male

with false alarm, Type I error

• False negative (FN) eqv. with

miss, Type II error
female
Ex: Performance
Confusion matrix
Classification Task:

• True positive (TP) eqv. with hit

• True negative (TN) eqv. with

correct rejection

• False positive (FP) eqv.

with false alarm, Type I error

• False negative (FN) eqv. with

miss, Type II error
Performance
Regression Task:

• Error measures

Example

+
1 3
!"# = ' , (() − ℎ1 (2 (() )
&
()*
 ML Applications

CARS HEALTH ROBOTICS

Self-driving cars will Genetic research, Autonomous robots

be in place by the medical assistance, to assist people.
end of the next and personalized
decade, 2030. treatments.
ML Social Ethics

1. The purpose specification

2. Limitations of data: bias

3. Limitations of data: privacy

4. Autonomous?

5. Education

6. People: Explainable AI