Lesson 2: How

computers learn from

data
Experience AI
 Starter activity
How tall is the The height of the
Eiffel tower? Eiffel tower is 330 m
or 1083 ft.

Is a ‘smart’ speaker an AI application?

Why?
Discuss with the person next to you.

You can use a ‘smart’ speaker (like Amazon Echo or Google

Home) to do things like answer a question, play a song, or
set an alarm.
 Objectives

Lesson 2: How computers learn from data

In this lesson, you will:
● Define machine learning’s relationship to artificial intelligence
● Name the three common approaches to machine learning
● Describe how classification can be solved using supervised learning
 Introduction

Two different approaches

Online streaming service

Spreadsheet that uses
that recommends films
formulas to work out the
based on users’ viewing
total cost of a holiday
habits
 Introduction

Two different approaches

Spreadsheet that
Rule-based uses formulas to
— work out the total
cost of a holiday

Online streaming
service that
Data-driven recommends films
— based on users’
viewing habits
 Introduction

Some uses of a ‘smart’ speaker

● Interpreting voice commands
How tall is the The height of the
● Answering questions Eiffel tower? Eiffel tower is 330
m or 1083 ft.
● Playing music
● Setting alarms

These are all things Which uses would

you use a ‘smart’ benefit from a data-
speaker to do. driven approach?
 Introduction

A data-driven approach
● Interpreting voice
There are many ways a person
commands
can phrase a command, and
● Answering questions people speak in different
● Playing music languages and have different
● Setting alarms accents.

In order to respond to as many

people as possible, a model is
beneficial.
 Introduction

A data-driven approach
● Interpreting voice
Answers are found on the web
commands
using a search engine.
● Answering questions
● Playing music Search engines might use a
model to provide high-quality
● Setting alarms
results. There might also be a
model to help pick which result is
the most relevant for the current
question.
 Introduction
Is a ‘smart’ speaker an AI
application?
Yes.

But just because something is an

AI application doesn’t mean How tall is the The height of the
every bit of it uses a data-driven Eiffel tower? Eiffel tower is 330 m
or 1083 ft.
approach.

When discussing AI applications,

you should be specific about
which use requires AI techniques.

Creators of ‘smart’ speakers

make use of AI to help
interpret voice commands
and answer questions.
 Activity 1

How do models ‘learn’?

If I wanted the model used in my
‘smart’ speaker to accurately
interpret commands from as
many people as possible… How tall is the
Eiffel tower?

What would I need to train

the model? The height of the
Eiffel tower is 330 m
or 1083 ft.

Data!
 Activity 1

What is machine learning?

Watch the video on YouTube

 Activity 1

How do models ‘learn’?

Machine learning is how the Machine learning:
models you have heard about are
created. Machine learning (ML) is a way
of building artificial intelligence
Reminder: A model is a systems using statistics,
representation of a real-world instead of by writing out the
context. rules in a program. Machine
learning systems are called
‘data-driven’, because they use
a lot of examples — data — to
work.
 Activity 1

How do models ‘learn’?

AI

Machine
The current Not all AI
state of AI learning applications
relies heavily use machine
on the learning.
techniques in
machine
learning.
 Activity 1

How do models ‘learn’?

Traditional programming is great
when we can define the exact
steps to solve a problem. Unlike traditional programming
where step-by-step instructions
Models help solve complex are written to solve a problem,
problems that require more than a machine learning systems are
rule-based approach. designed to make predictions
based on lots of data.
 Activity 1

How do models ‘learn’?

The outputs from a machine
learning model are predictions
based on the data used to create Unlike traditional programming
it. There is always some where step-by-step instructions
uncertainty in the output. are written to solve a problem,
machine learning systems are
Where have you seen uncertainty designed to make predictions
in a model’s output before? based on lots of data.
 Activity 1
Would you use machine
learning?
Not all problems require a machine learning model.

On your worksheet, there are three desired uses for an application.

Label the uses with the approach you think would be appropriate,
either traditional instructions or a machine learning model.
 Activity 1
Would you use machine
learning?

Calculatin
Forecastin Identifyin
g the
g air g animals
distance
pollution in a video
to a star
 Activity 1
Would you use machine
learning?

Calculatin
Forecastin Identifyin
g the
g air g animals
distance
pollution in a video
to a star

Probably Could use Could use

instructio an ML an ML
ns model model
 Activity 1

What data would you need?

In order to create a machine
learning model, you need data! Identifyin
g animals
Pick one of the two uses of in a video
machine learning models. What
data would you need to create a
model?
Forecastin
Write your answers on your g air
worksheet. pollution
 Activity 2

Types of machine
learning

Watch the video on YouTube

 Activity 2

Types of machine learning

Which type of machine learning do the
following statements apply to?

Supervise Unsupervise Reinforceme

d learning d learning nt learning

1 2 3
 Activity 2

Types of machine learning

Uses an agent that has direct access to the

environment

Supervise Unsupervise Reinforceme

d learning d learning nt learning

1 2 3
 Activity 2

Types of machine learning

Uses an agent that has direct access to the

environment
Reinforcement
learning
Supervise Unsupervise
d learning d learning

1 2 3
 Activity 2

Types of machine learning

Training requires example data

Supervise Unsupervise Reinforceme

d learning d learning nt learning

1 2 3
 Activity 2

Types of machine learning

Training requires example data

Supervised
learning Unsupervise Reinforceme
d learning nt learning

1 2 3
 Activity 2

Types of machine learning

Grouping data that is similar

Supervise Unsupervise Reinforceme

d learning d learning nt learning

1 2 3
 Activity 2

Types of machine learning

Grouping data that is similar

Unsupervised
Supervise learning Reinforceme
d learning nt learning

1 2 3
 Activity 2

Types of machine learning

Learning by trial and error

Supervise Unsupervise Reinforceme

d learning d learning nt learning

1 2 3
 Activity 2

Types of machine learning

Learning by trial and error

Reinforcement
learning
Supervise Unsupervise
d learning d learning

1 2 3
 Activity 2

Types of machine learning

Requires lots of training data

Supervise Unsupervise Reinforceme

d learning d learning nt learning

1 2 3
 Activity 2

Types of machine learning

Requires lots of training data

Supervised
learning Unsupervise Reinforceme
d learning nt learning

1 2 3
 Activity 2

Types of machine
learning
Supervised Unsupervised Reinforcement
learning learning learning

Needs data that has Clusters similar data Learns by trial and
been preprocessed together error
by a human

Requires a lot of Gives you ideas for Model needs time to

data to train how to group the get good at its tasks
data
1. Denetimli Öğrenme (Supervised
Learning)
● Tanım: Etiketli veri ile çalışan, giriş-çıkış eşleşmeleri üzerinden tahmin yapan
algoritmalardır.

● Amaç: Yeni gelen veriler için doğru tahmin yapmak.

● Nasıl çalışır: Giriş verisi (X) ve doğru çıkış (Y) ile model eğitilir.
● Örnekler:
● • Spam e-posta tespiti
● • Resimde kedi-köpek ayrımı
● • Ev fiyatı tahmini
● Kullanılan Algoritmalar:
● • Karar ağaçları, Lojistik regresyon, Yapay sinir ağları
2. Denetimsiz Öğrenme (Unsupervised
Learning)
● Tanım: Etiketlenmemiş verilerle çalışır. Veri içindeki gizli örüntüleri
keşfetmeye çalışır.

● Amaç: Veriyi gruplamak veya yapısını anlamak.

● Nasıl çalışır: Sadece giriş verisi (X) vardır, çıkış (Y) yoktur.
● Örnekler:
● • Müşteri segmentasyonu
● • Anomali (sapma) tespiti
● • Konu modelleme
● Kullanılan Algoritmalar:
● • K-Means, Hiyerarşik kümeleme, PCA
3. Pekiştirmeli Öğrenme (Reinforcement
Learning)
● Tanım: Ajan bir ortamda hareket eder ve yaptığı eylemler karşılığında ödül veya
ceza alır.

● Amaç: Uzun vadede maksimum ödülü getiren stratejiyi öğrenmek.

● Nasıl çalışır: Ajan her adımda karar verir, ortamdan ödül alır ve gelecekteki
kararlarını bu ödüllere göre iyileştirir.
● Örnekler:
● • Oyun oynayan yapay zekâ
● • Otonom araçlar
● • Robotların yol bulması
● Kullanılan Algoritmalar:
● • Q-learning, DQN, Policy Gradient
Öğrenme Türlerinin Karşılaştırması
● Etiketli Veri: ● • Supervised: Tahmin, sınıflandırma
● • Supervised: Evet ● • Unsupervised: Segmentasyon,
● • Unsupervised: Hayır desen bulma
● • Reinforcement: Hayır ● • Reinforcement: Karar verme,
oyunlar
● Çıkış Bilgisi:
● • Supervised: Bilinir
● • Unsupervised: Bilinmez
● • Reinforcement: Ödül/ceza

● Kullanım Alanları:
 Activity 3

Classification
Classification is a way of using supervised learning to
categorise data.

A classification model assigns data to a class by applying labels.

Apple Banana Orange

 Activity 3

Classification
Classification models must be trained with example data that
already has labels assigned by a human.

What labels would you apply to these images?

 Activity 3

Classification
Classification models must be trained with example data that
already has labels assigned by a human.

What labels would you apply to these images?

Orange Banana Apple

 Activity 3

Classification
The more training data you use, the more accurate the
model will be. Banana

Apple Orange

Classificati
on model
 Activity 3

Classification
Once the model is trained, new data can be fed into the model
and it will produce a prediction of which class the data belongs
to.
New
data

Prediction:
Classificati
on model Orange
Confidence 92%
 Activity 3

Classification
Class Label

A category that the Applied to a single piece of

data can be assigned to data to indicate which class
it belongs to
Apple

Apple
 Activity 3

Classification —
example 1
Classification can be used to
determine whether reviews or Beautiful Colour
Reviewed in the United Kingdom 🇬🇧 on 22 September 2022

comments are positive or Size Name: 128GB | Colour: Alpine Green | Configuration: Without AppleCare+ | Verified
Purchase
Very happy with this phone.
Is my first iPhone, as I always had Windows Phone and then Android, but now I don’t think I

negative. could go back. Apple is super pricey so I think I will keep this phone for many years, but
generally, I’m happy and the colour is stunning.
Great battery life, great photos, and videos, very smooth.

Helpful Report abuse

The model is trained with text

data. Waste of money
Reviewed in the United Kingdom 🇬🇧 on 20 September 2022
Size Name: 128GB | Colour: Graphite | Configuration: Without AppleCare+ | Verified
Purchase
Disappointed after two weeks only battery drain was so big i had to charge it 2-3 times a day
waste of money

This is called sentiment Helpful Report abuse

analysis.
 Activity 3

Classification —
example 1
Sentiment analysis
Beautiful Colour
Reviewed in the United Kingdom 🇬🇧 on 22 September 2022

Classes: Positive and Negative Size Name: 128GB | Colour: Alpine Green | Configuration: Without AppleCare+ | Verified
Purchase
Very happy with this phone.
Is my first iPhone, as I always had Windows Phone and then Android, but now I don’t think I
could go back. Apple is super pricey so I think I will keep this phone for many years, but
generally, I’m happy and the colour is stunning.
Great battery life, great photos, and videos, very smooth.

Benefits: Can process a larger Helpful Report abuse

number of reviews to get an

accurate picture of how well your Waste of money
Reviewed in the United Kingdom 🇬🇧 on 20 September 2022
Size Name: 128GB | Colour: Graphite | Configuration: Without AppleCare+ | Verified

product or service has been Purchase

Disappointed after two weeks only battery drain was so big i had to charge it 2-3 times a day
waste of money

received Helpful Report abuse

 Activity 3

Classification —
example 2
Computer vision

Using a model to find a range of

objects in a photo or video.

The model is trained with images

of individual objects and of
multiple objects, like this image.
 Activity 3

Classification —
example 2
Shoe 1
● Confidence: 100%

Person 2
● Confidence: 99%

Dog 3
● Confidence: 98%

Bicycle 4
● Confidence: 95%

Wheel 5
● Confidence: 81%
 Activity 3

Classification —
example 2
Computer vision

Classes: Shoe, person, dog,

bicycle, wheel, etc.

Benefits: Can identify multiple

objects in an image — useful for
technological advancements such
as driverless cars
 Activity 3

Data types for

classification
You can use almost any type of Beautiful Colour
Reviewed in the United Kingdom 🇬🇧 on 22 September 2022
Size Name: 128GB | Colour: Alpine Green | Configuration: Without AppleCare+ | Verified Purchase

data to train a classifier. Very happy with this phone.

Is my first iPhone, as I always had Windows Phone and then Android, but now I don’t think I could
go back. Apple is super pricey so I think I will keep this phone for many years, but generally, I’m
happy and the colour is stunning.
Great battery life, great photos, and videos, very smooth.

Helpful Report abuse

On your worksheet, look at the Waste of money

Reviewed in the United Kingdom 🇬🇧 on 20 September 2022
Size Name: 128GB | Colour: Graphite | Configuration: Without AppleCare+ | Verified Purchase

example classification problems.

Disappointed after two weeks only battery drain was so big i had to charge it 2-3 times a day
waste of money

Helpful Report abuse

● What type of data would you

use?
● What could some of the
classes be?
 Activity 3

Data types for

classification

Assigning Identifyin Languag

a genre g spam e
to music emails detection

What type of data would you use?

What could some of the classes
 Plenary

Classification — your
turn
DeepMind want you to help
them with their project of
tracking animals in the Serengeti.

The Serengeti is a national park

in East Africa. There are many
endangered animals there.
Watch the video on YouTube
They are using classification to
find and track animals in order to
protect them.
 Plenary

Classification — your
turn Classes They need your help to pre-
label some images to help
cheetah hyena gazelle
train their classification
model.

What labels would you

apply to the following
zebra elephant lion images?
 Plenary

Classification — your
turn Classes
cheetah hyena gazelle

zebra elephant lion

 Plenary

Classification — your
turn Classes
cheetah hyena gazelle

zebra elephant lion

 Plenary

Classification — your
turn Classes
cheetah hyena gazelle

zebra elephant lion

 Plenary

Classification — your
turn Classes
cheetah hyena gazelle

zebra elephant lion

 Summary

Next lesson
In this lesson, you… Next lesson, you will…

Described which parts of a Explore the importance of

system use AI data

Explored the role of data in AI Learn about training and test

applications data

Defined ‘machine learning’ Describe ways to identify and

prevent bias in your ML
Used classification to
models
categorise data