Interactive learning material

Machine Learning

Introduction to the Conceptual Framework of Artificial Intelligence

Author: Zsolt LengyeliTStudy Hungary Training and Research Centre
Reviewer: Mária HartyányiiTStudy Hungary Training and Research Centre

The three building blocks of machine learning: data, models and learning paradigms — The building blocks of machine learning

Learning goal

The rapid rise of artificial intelligence has triggered extreme reactions, ranging from boundless enthusiasm to the fear that it will soon take our jobs. This learning material reviews the theoretical background needed for a realistic assessment of AI and clarifies the basic concepts essential for informed use.

By the end of the learning material, you will find answers to these questions

How does a computer learn and communicate?
What do we mean by machine learning?
What does it mean that AI systems do not operate deterministically?
Is artificial intelligence a machine version of human intelligence?
Can artificial intelligence be smarter than humans?

Chapter 1 · 1 / 2

The Idea of a Machine That Can Learn

The famous mathematician Alan Turing put it this way in a 1947 lecture to the London Mathematical Society:

“What we want is a machine that can learn from experience.”Alan Turing, 1947

The computers available at the time did not make it possible to build such a program, and it took more than seventy years to get closer to Turing's vision.

Following the 1956 Dartmouth conference, researchers gradually developed the mathematical models and algorithms required for machine learning. At the same time, huge volumes of digital data became available, together with the computing capacity needed to process them.

1947Turing proposes the idea of a machine that learns from experience.

1956The Dartmouth conference gives momentum to AI research.

2022ChatGPT brings language-based AI to widespread public attention.

Chapter 1 · 2 / 2

When a Machine Responds in Human Language

ChatGPT was launched on 30 November 2022. The number of registered users reached one million within five days — compared with three and a half years for Netflix.

Its interface let anyone experience communication with a computer through natural human language. Its quick, coherent and convincing answers simultaneously sparked enthusiasm, surprise and concern.

A human face and a digital face representing artificial intelligence — Human-like depictions of AI influence what we assume about its capabilities.

An older woman speaks on the phone while artificial intelligence is represented at the other end — When “someone” speaks in a perfectly human voice

Chapter 2

Is the Machine Really Learning?

When we say that a computer “learns”, we do not mean the same process by which a human being, especially a young child, learns. Science still cannot fully describe the complex processes of human learning.

Machine learning: the collective term for algorithms and methods that learn from data. A computer recognises patterns from data and examples, then uses them to generate predictions or decision recommendations.

The “recipe” has three ingredients: digitally stored data, a learning algorithm that specifies how learning takes place, and a model in which the learned relationships are stored.

01

Data

Examples and observations from which the system learns patterns.

02

Algorithm

The method that determines how learning takes place.

03

Model

The result of training, which generates inferences from new data.

Chapter 3 · 1 / 4

Supervised Learning

In supervised learning, the machine learns from examples in which the correct answers are provided along with the input data. An example could be millions of photographs labelled with the names of the objects shown in them.

Imagine training a model to decide whether a picture shows an apple or a banana. The predefined correct answers are called labels, and these examples are known as labelled data.

From the examples, the model learns visual features such as shape, colour and other patterns. It can then classify images it has never seen before — with a certain probability.

Supervised learning illustrated with labelled pictures of dogs and cats — Examples with correct labels guide the learning process.

Chapter 3 · 2 / 4

From Training Data to a Working Model

Large amounts of data are used for training; this is called the training database or training dataset. The resulting model can identify new, unknown data based on the relationships it recognised during training.

An important limitation: the model can only draw conclusions from what it learned from the training data. It has no independent knowledge or experience beyond that data.

A flowchart of input data, a learning algorithm and a model — Machine learning: building a model from labelled examples

Speech recognition icon — Speech recognition

Text generation icon — Text generation and translation

Chapter 3 · 3 / 4

Unsupervised Learning

In unsupervised learning there is no predefined correct answer: the machine receives no labels. The algorithm independently discovers hidden patterns, clusters, similarities and differences in the dataset.

These methods support recommendation systems: they identify interests from user behaviour and recommend personalised films, music or products.

Fruit grouped automatically according to similarity — The system searches for structure in the data without labels.

Chapter 3 · 4 / 4

Reinforcement Learning

In reinforcement learning, a virtual “agent” interacts with its environment and tries to make optimal decisions. Through rewards or penalties, it learns which actions lead to the best outcome.

The cycle of agent, environment, action and reward — The agent learns from the consequences of its actions.

In 1997, IBM's Deep Blue defeated world chess champion Garry Kasparov using traditional programming, enormous computing power and a strong chess-evaluation system. The game of Go, however, has too many possible positions to enumerate completely.

DeepMind's AlphaGo used reinforcement learning and practised against itself until it could defeat Lee Sedol, one of the world's best players, in a match series.

AlphaGo playing against Lee Sedol — AlphaGo and Lee Sedol during their match (Source: Google Arts & Culture)

Chapter 4

Transfer Learning and Application Areas

Transfer learning

A completed, well-functioning model can be developed further by continuing its training in a narrower, specialised area — for example, by training a general language model on carefully selected educational sources.

Supervised learning

Classification and prediction

Object recognition, spam filtering, credit-risk assessment and support for medical diagnostics.

A schematic illustration of classification

Unsupervised learning

Discovering hidden structure

Customer groups, anomalous events and the simplification of complex data.

Reinforcement learning

Sequences of decisions

Robot control, game programs, transport optimisation and automated control tasks.

Chapter 5

The Resources and Costs of Machine Learning

The resource requirements of machine learning methods depend heavily on the amount of data, the complexity of the model and the chosen learning method.

Supervised learning: collecting and labelling training data can be expensive.
Unsupervised learning: storing and processing large datasets can require substantial computing capacity.
Reinforcement learning: many trials, often in a simulated environment, are needed.

Cinderella sorts different seeds into groups — Like Cinderella sorting seeds, unsupervised learning finds natural groups in data without labels.

Takeaways for teachers

What AI Can — and Cannot — Do

“Artificial intelligence does not think, it is talking.”Vilmos Csányi, 2026

Knowledge check

Check What You Have Learned

1What is a machine learning model? The result of training, which uses learned patterns to make inferences from new data. A set of rules that is infallible in every situation. A digital copy of the human brain.

2Which method uses examples supplied with correct answers? Supervised learning Unsupervised learning Reinforcement learning

3What characterises unsupervised learning? It searches for patterns and groups without labels. It learns exclusively from rewards. It always gives the same answer.

4How does a reinforcement learning system learn? It evaluates its actions using rewards and penalties. It memorises every possible answer. It only uses labelled images.

5Which statement is true? AI responses are based on learned patterns and therefore require human review. A natural voice proves that AI is conscious. A model's knowledge is independent of its training data.

Contributors and attachments

Credits and Copyright

Author: Zsolt LengyeliTStudy Hungary Training and Research Centre
Reviewer: Mária HartyányiiTStudy Hungary Training and Research Centre
Project: 2025-1-LT01-KA220-VET-000362456

Attachments

Download the English source materialM1-T1-MLC1-TEXT-EN.pdf Download the practical exerciseM1-T1-MLC1-EXC-EN.pdf

Creative Commons Attribution-ShareAlike emblem

This work may be reused and adapted under the terms of the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0).

Funded by the European Union. Views and opinions expressed are those of the author(s) only and do not necessarily reflect those of the European Union or the National Agency.