Machine Learning Explained: A Beginner's Guide

Learning from examples instead of following rules

What changes when a program is not given step-by-step instructions, but works them out from examples? Everything, or nearly. That is the idea behind machine learning, a technique within artificial intelligence.

In plain words, and according to MIT Technology Review (17 November 2018), machine-learning algorithms use statistics to find patterns in huge amounts of data. That data can be numbers, words, images or clicks: anything that can be stored digitally.

From Samuel’s draughts to Mitchell’s definition

The term has a documented history. In July 1959, Arthur L. Samuel published “Some Studies in Machine Learning Using the Game of Checkers” in the IBM Journal (vol. 3, no. 3). Let’s go to the paper: its abstract states that enough work has been done to verify that a computer can be programmed to learn to play draughts better than the person who wrote the program.

According to the transcript of the first CS229 lecture at Stanford, the program played against itself and learned which board positions tend to be associated with wins and which with losses. The same transcript attributes to Samuel an informal definition: the field that gives computers the ability to learn without being explicitly programmed. Note the nuance: this is the definition the lecture attributes to him, not a verbatim quotation from the paper.

The transcript also gives Tom Mitchell’s formal definition. Put simply: a program learns from experience E with respect to a task T and a performance measure P if its performance at T, as measured by P, improves with E.

An example of my own, different from the draughts one: a filter that classifies emails as spam or not spam. The task (T) is classifying emails; the experience (E) is emails that someone has already flagged; the measure (P) could be the percentage of emails classified correctly. If that percentage rises as the filter sees more examples, it is learning in Mitchell’s sense.

In other words, the definition forces you to ask how the improvement is measured. It is worth keeping in mind: without P, “learning” means nothing.

The three ways of learning: supervised, unsupervised and reinforcement

According to MIT Technology Review, machine learning (and deep learning) comes in three flavours.

Supervised. The data is labelled to tell the machine which patterns to look for. According to the article, this is the most widespread. DeepMind (25 June 2019) specifies that the training signal is the correct labels, for example the label of an image.
Unsupervised. The data has no labels and the machine looks for whatever patterns it can find. DeepMind presents it as a paradigm in which the agent learns from the data it observes, without a specific task.
Reinforcement. The system learns by trial and error to reach a clear goal, with rewards or penalties. According to MIT Technology Review, it is the basis of Google’s AlphaGo. DeepMind gives another example: a high score in an Atari game.

As early as 1959, Samuel distinguished two general approaches. One, the “neural-net approach”, seeks to induce learned behaviour in a randomly connected network by means of a reward-and-punishment routine. The other, far more efficient according to the text, uses a highly organised network designed to learn only certain specific things. The idea of reward and punishment was already there, although the sources consulted do not allow us to claim a direct lineage to today’s reinforcement learning.

Which methods exist and how we check they work

The scikit-learn user guide (version 1.9.1; 1.9.0 was released in June 2026, according to the project’s website) offers a map of methods.

Supervised: linear models, support vector machines, nearest neighbours, decision trees, Naive Bayes and neural networks.
Unsupervised: Gaussian mixtures, clustering with DBSCAN and others, dimensionality reduction and outlier detection.

The sources consulted do not detail how each family works, so I do not unpack them here. For that, see the article on the most common types of machine-learning algorithm.

What that documentation does show is that the work is not just choosing an algorithm. It includes sections on data preprocessing and on evaluation, with metrics for classification, regression and clustering. Preparing the data and measuring results is part of the craft, and it ties back to Mitchell: the measure P is no ornament.

The limits: who decides what the machine learns

According to DeepMind, the advances of the last decade in image recognition, self-driving cars and games such as Go have been achieved largely by training deep neural networks with supervised or reinforcement learning. But the same text points out that both paradigms require training signals designed by humans, so the limits of the learning are set by those who do the training.

This means that the labels, the rewards and the chosen data are all decisions made by people. What the text does not claim is that this is enough: it presents unsupervised learning as a way of learning from what is observed without a specific task.

What we still do not know, at least with these sources: how much data is needed, which typical errors arise, or how biases are measured. The sources consulted do not go into detail, and I would rather not fill the gap with figures from elsewhere. If you are curious about the machinery behind many of these models, you can go on to how neural networks work.

Where it fits in European regulation

Regulation (EU) 2024/1689 on Artificial Intelligence (the AI Act) is, according to AESIA (the Spanish Agency for the Supervision of Artificial Intelligence), the world’s first legal framework on AI. It entered into force on 1 August 2024 and aims to make AI systems safe, ethical and trustworthy, with a high level of protection of health, safety and fundamental rights.

According to the AESIA practical guide, the Regulation defines an AI system as a machine-based system designed to operate with varying levels of autonomy and which may exhibit adaptiveness after deployment. For explicit or implicit objectives, it infers from the input it receives how to generate outputs (predictions, content, recommendations or decisions) that can influence physical or virtual environments. The guide takes as its reference the version published by the Council of the European Union on 13 June 2024 and notes that the European Commission has published guidelines to clarify that definition.

Opinion: “inferring from the input data” sounds very much like what a trained model does. That is my own reading, not a legal equivalence: the guide does not say that every definition of machine learning matches the legal one.

My reading and where to go next

Opinion: a definition of machine learning is worth only as much as its evaluation. Whenever someone says a system “learns”, I would ask three things: what was it given, what was measured and who decided. With those three answers you understand almost everything else.

If you want to see how it works on the inside, the natural next step is the guide to what deep learning is, where Mar explains it from the mechanics.

Sources

Image: geralt on Pixabay

Translated and adapted by Ollie.

What is machine learning? A beginner’s guide