Learning Systems (DT8008)

Introduction to
Machine Learning
Dr. Mohamed-Rafik Bouguelia
[email protected]

Halmstad University
 What is
Machine Learning ?
What is Machine Learning (ML) ?
• Machine Learning: Field of study that gives computers the
ability to learn without being explicitly programmed with rules.

3
What is Machine Learning (ML) ?
Definition:
• A computer program is said to learn from experience E with respect to
some task T and performance measure P if its performance on T, as
measured by P, improves with experience E.

Example:

Suppose your email

program watches
which emails you do
or do not mark as
spam, and based on
that learns how to
better filter spam.
What is the task T in
this setting ?

4
What is Machine Learning (ML) ?
Definition:
• A computer program is said to learn from experience E with respect to
some task T and performance measure P if its performance on T, as
measured by P, improves with experience E.

Example:
1. Classifying emails as spam or not spam.
Suppose your email
program watches 2. Watching you label emails as spam or not spam.
which emails you do
or do not mark as
spam, and based on 3. The number (or fraction) of emails correctly
that learns how to classified as spam/not spam.
better filter spam.
What is the task T in
this setting ? 4. None of the above—this is not a machine
learning problem.

5
What is Machine Learning (ML) ?
Definition:
• A computer program is said to learn from experience E with respect to
some task T and performance measure P if its performance on T, as
measured by P, improves with experience E.

Example:
T 1. Classifying emails as spam or not spam.
Suppose your email
program watches E 2. Watching you label emails as spam or not spam.
which emails you do
or do not mark as
spam, and based on P 3. The number (or fraction) of emails correctly
that learns how to classified as spam/not spam.
better filter spam.
What is the task T in
this setting ? 4. None of the above—this is not a machine
learning problem.

6
 What is Machine Learning (ML) ?
• Usual programming
Data
Computer Output
Rules (program)

• (Supervised) Machine learning

Data Training
Training dataset
Computer ML Model (Learning)
(data + output)
Output

New Data Model Output Predicting

• Machine learning algorithms build a model from the training data, then uses this model to
make predictions or decisions without being explicitly programmed to perform the task.
What is Machine Learning (ML) ?
Example:
The data consist of
images …

The output consists of

labels (Cat, Dog or Lion)
• (Supervised) Machine learning

Data Training
Computer ML Model (Learning)
Output

New Data Model Output Predicting

Dog (predicted label)

Cat 5%, Dog 65%, lion 30%
 What is Machine Learning (ML) ?

Machine learning types:

- Supervised learning
- Unsupervised learning
- Others
- Reinforcement learning
- Semi-supervised learning
- Active learning
- etc.

Supervised ML: Unsupervised ML:

Training data includes desired outputs. Training data does not include outputs.
Anomaly detection

9
Question
• You want to do some task …
– e.g. predicting if an email is a spam or not.

• Why would you need machine learning?

• Why don’t you just explicitly program/write rules to

perform the task (without ML) ?
– e.g. if the email is from an unknown sender and contains
keywords such as
• "send x usd", "only for you", "invest now", "your computer is
compromised" …
– then it’s a spam

10
Example (self-driving car)
Consider the following problem:
• You have a camera on your car that
periodically captures images of the road
and send them to your app.
• You want your app to recognize what is
present on each image (pedestrians,
bikes, other cars, etc …)

Question:
• Why do we need machine learning for
this? Why don’t we just explicitly
program/write rules that allows us to
recognize what the image contains?

11
Example (self-driving car)
Consider the following problem:
• You have a camera on your car that
periodically captures images of the road
and send them to your app.
• You want your app to recognize what is
present on each image (pedestrians,
bikes, other cars, etc …)

Question:
• Why do we need machine learning for
this? Why don’t we just explicitly
program/write rules that allows us to
recognize what the image contains?

12
 Introduction to
Supervised Machine Learning

Regression problems
Introduction to Supervised Learning
• Housing price prediction (regression)

14
Introduction to Supervised Learning
• Housing price prediction (regression)
• Suppose that you want to sell a house of size 750 feet² and want to know
how much you can get for this house, i.e. predict its price.
• How can a learning algorithm help you?

750
15
Introduction to Supervised Learning
• Housing price prediction (regression)
• you can fit a straight line to the data, and predict the price of the house.

150

750
16
Introduction to Supervised Learning
• Housing price prediction (regression)
• you can fit a straight line to the data, and predict the price of the house.
• or maybe its better to fit a quadratic function (2nd order polynomial).

750
17
Introduction to Supervised Learning
• Housing price prediction (regression)
• you can fit a straight line to the data, and predict the price of the house.
• or maybe its better to fit a quadratic function (2nd order polynomial).

How to decide
which model to
choose for this
dataset ?
We will see this
later in the
course when we
talk about model
selection …

18
 Introduction to Supervised Learning
• Housing price prediction (regression)
 This is an example of a supervised learning algorithm:
 The right answers (here, the prices) are given in the training dataset.
 More specifically, this example was a regression problem:
 Predicting a continuous valued output (price).
Continuous
values

19
Introduction to Supervised Learning
Other regression examples

20
 Introduction to
Supervised Machine Learning

Classification problems
Introduction to Supervised Learning
• Breast cancer malignant/benign (classification)

(feature 1)

We only have discrete output values (in this example: 1 or 0)

22
 Introduction to Supervised Learning
• Breast cancer malignant/benign (classification)
We don’t need the second axis if we simply visualize the
classes with different colors or shapes.

Two
axis

(feature 1)
Malignant ?
Only
one
axis
(feature 1)
23
Introduction to Supervised Learning
• Breast cancer malignant/benign (classification)
– The patients data can be characterized by more than one feature
• e.g.Tumor size and Age …
Malignant ?

(feature 2)

(feature 1) 24
Introduction to Supervised Learning
• Breast cancer malignant/benign (classification)
– Suppose that you get a new patient who has some tumor size s and age a, and you want
to predict if it is malignant or benign. How can a learning algorithm help you?

Malignant ?

a ?

(feature 2)

s
(feature 1) 25
Introduction to Supervised Learning
• Breast cancer malignant/benign (classification)
– you can fit a linear model to the training data, then predict the class of the new patient.

Malignant ?

a ?
So for patient (s, a),
(feature 2) we would predict the
class “benign”.

s
(feature 1) 26
Introduction to Supervised Learning
• Breast cancer malignant/benign (classification)
– you can fit a linear model to the training data, then predict the class of the new patient
– or you can fit a non-linear model to the training data …

Malignant ?

a ?
So for patient (s, a),
(feature 2) we would predict the
class “benign”.

s
(feature 1) 27
Introduction to Supervised Learning
• Breast cancer malignant/benign (classification)
 Again, this is an example of a supervised learning
algorithm:
 The right answers (here, the classes malignant /
benign) are given with the training dataset.
 i.e. for each patient (data-point) in the training
dataset, we know if he is has a malignant or benign
cancer.

 However, this example was a classification problem:

 Predicting a discrete valued output (malignant /
benign).

Note: In this example we had two features (age, size), but we will see ML
algorithms that can easily deal with a much larger number of features …

28
Introduction to Supervised Learning
Other classification examples • Classification is about learning
decision boundaries, and predicting
the “class” of new data-point.
speed

In 3-dimensions, this linear model is a “plan”.

In more than 3-dimensions, it’s a “hyperplan”.

size

29
 Introduction to
Supervised Machine Learning

Difference between
Regression and Classification
 Difference between Regression and Classification

Regression:
• The output (i.e. target variable) is continuous. It consist of

output
real values.
– predicting the price of houses (in SEK)
– predicting the power consumption (in kW)
– predicting how much healthy is the patient (e.g. ∈ [0, 1])
– etc. feature 1

Classification:
• The output (i.e. target variable) is discrete. It consists of

feature 2
class A
classes (or categories). class B
– predicting if an image contains a cat or a dog
– predicting customer categories
– good/bad, healthy/sick, red/green/blue, A/B/C/D, 0/1/2
… feature 1
– etc.
31
Difference between Regression and Classification

• You’re running a company, and you want to develop machine learning

algorithms to address each of two following problems:

• Problem 1:
– You have a large inventory of identical items.You want to predict how many of
these items will sell over the next 3 months.

• Problem 2:
– You’d like software to examine individual customer accounts, and for each
account decide if it has been hacked/compromised.

• Should you treat these as classification or as regression problems?

– Treat both as classification problems.
– Treat problem 1 as a classification problem, problem 2 as a regression problem.
– Treat problem 1 as a regression problem, problem 2 as a classification problem.
– Treat both as regression problems.

32
Difference between Regression and Classification

• You’re running a company, and you want to develop learning algorithms to

address each of two following problems.

• Problem 1: The output is the number of items. Time is a feature here.

– You have a large inventory of identical items.You want to predict how many of
these items will sell over the next 3 months.

• Problem 2: The output consists of two classes: hacked / not hacked

– You’d like software to examine individual customer accounts, and for each
account decide if it has been hacked/compromised.

• Should you treat these as classification or as regression problems?

– Treat both as classification problems.
– Treat problem 1 as a classification problem, problem 2 as a regression problem.
– Treat problem 1 as a regression problem, problem 2 as a classification problem.
– Treat both as regression problems.

33
 Introduction to
Unsupervised Learning
Introduction to Unsupervised Learning
In supervised learning (e.g. classification) we have a labeled training dataset:

So, for each data-point ∈ 𝑅𝑅 2 , we have the corresponding class-label ∈

35
Introduction to Unsupervised Learning
In unsupervised learning (e.g. clustering) we have an unlabeled training dataset:

In clustering, we want to explore the data

2
We only have data-points ∈ 𝑅𝑅 to find some intrinsic groups (clusters) in
it. The clusters are not known beforehand.

36
Introduction to Unsupervised Learning
Some applications of clustering

Automatically
grouping together the
stories (news articles
on the Web) that talk
about the same topic.

37
Introduction to Unsupervised Learning
Some applications of clustering

38
Introduction to Unsupervised Learning
Some applications of clustering

• DNS Microarray data.

• Colors here corresponds to how much

individuals do or do not have a certain gene.

39
Introduction to Unsupervised Learning
Some applications of clustering

• DNS Microarray data.

• Colors here corresponds to how much

individuals do or do not have a certain gene.

• Run a clustering algorithm to group individuals

into different groups/types of people.

40
Introduction to Unsupervised Learning
Some applications of clustering

41
Introduction to Unsupervised Learning

42
Introduction to Unsupervised Learning

43
Course contents and format
Course contents and format
• Week 4 (Basics)
– Lecture 1.1 Introduction to machine learning (this lecture).
– Lecture 1.2 Basics, prerequisite, and review of important notions.
– Lab 1: Hands-on Python for ML
• Week 5 (Regression)
– Lecture 2.1 Linear Regression.
– Lecture 2.2 Nonlinear Regression (KNN and Kernel regression)
– Lab 2: Implementing linear regression (with/without gradient descent) + Kernel regression.
• Week 6 (Classification)
– Lecture 3.1 Classification using Logistic Regression.
– Lecture 3.2 Nonlinear Classification (Polynomial features, KNN, DTrees, …)
– Lab 3: Implementing logistic regression + KNN.
• Week 7 (Generalization)
– Lecture 4.1 Overfitting and Regularization.
– Lecture 4.2 Ensemble Methods (Random Forest).
– Lab 4: Re-implementing LinReg and LogisticReg with Regularization + Implementing Random Forest.
• Week 8 (SVM)
– Lecture 5 Support Vector Machines.
– Lab 5: Using SVM (Linear + with Kernel Trick) for Spam Classification.
• Week 9 (ANN)
– Lecture 6.1 Artificial Neural Networks (ANN).
– Lecture 6.2 Artificial Neural Networks (ANN) – Continuation.
– Lab 6: Implementing a simple ANN.
• Week 10 (Unsupervised Learning)
– Lecture 7.1 Dimensionality Reduction (using Principal Components Analysis)
– Lecture 7.2 Clustering
– Lab 7: Implementing PCA + K-means clustering.
• Week 11 (Presentations)
45
– Seminars where your present your projects …
Course contents and format
1. Written examination (3 credits)
– Mainly based on the contents of lectures.
– and some content related to the Labs.

2. Practical Projects and Labs (3 credits)

– The weekly Labs (jupyter notebooks).
• The Labs are to be done individually.
– Written report about the final project (to submit before week 11).
• The final project can be done in a group of one or two students (maximum).

3. Seminars (1.5 credits)

– Oral presentation of the final project (on week 11)
• The presentation (in a group of one or two students) should take 20 to 25 minutes max.
• The slides should show the project results achieved so far as well as a state-of-the-art
section which refers to research articles/papers related to your project (use Google
Scholar to find relevant papers).

46
Course contents and format
• The report should be about 7 to 10 pages including figures
and tables). It can be structured as follows:
1. Introduction
• Brief presentation of problem, 1 page.
2. State-of-the-art
• Brief description of research papers doing work related to your
project, 1 page.
3. Methodology
• Brief listing of methods used, 1 page.
4. Data
• Presentation of your dataset with important observations, 1-2 pages.
5. Results and their interpretation (3-5 pages).
6. Discussion
• Conclusions about your results and comparison to other
researchers' results, 1 page.
47
Course contents and format
• Regarding Labs:
– There is a Lab to do on each week (total of 7 Labs).
– You have to start working on each Lab soon after the corresponding
lecture (i.e. before the Lab session) and prepare questions for the Lab
assistant who will help you during the Lab session.
– The deadlines to submit each Lab are on Blackboard.
– Submit your Lab solutions as jupyter notebooks to the Lab assistants:
Reza [email protected] and Yuantao [email protected] and add
Rafik (as cc) [email protected]

• Regarding Projects:
– You have to submit your written report before week 11 to mohamed-
[email protected]

48