Machine learning: what it is and how a machine learns

The reversal: examples instead of rules

For decades, programming meant one thing: writing rules. If the email contains this word, flag it as spam; if the customer spends above this amount, apply the discount. It works as long as the rules can be written. But try writing the rules for recognizing a cat in a photo: pointy ears? So do foxes. Fur? Not sphynx cats. You give up fast.

Machine learning flips the table: instead of writing rules, you show the software thousands of examples (photos labeled “cat” or “not cat”) and let it dig out the patterns. The result of that studying is called a model. As Google puts it, ML is “training a piece of software, called a model, to make useful predictions or generate content from data”.

The idea has a precise birthday: in 1959 Arthur Samuel, at IBM, coined the term while describing his program that got better at checkers game after game, eventually beating skilled players.

The three ways of learning (plus one)

Supervised: studying from solved exams

The most common method: every example comes with the right answer attached. A thousand emails labeled “spam” or “not spam”, ten thousand houses with their sale price. The model learns the relationship between input and answer, like a student preparing from past years’ solved exams. The two classic applications: regression (predicting a number, like a house price) and classification (picking a category, like spam or not).

Unsupervised: finding order without instructions

Here the examples carry no labels. The model receives raw data and hunts for structure and similarity on its own: clustering groups customers with similar habits without anyone defining the groups in advance. Making sense of what emerges is then a human job.

Reinforcement: learning by failing

The system acts in an environment and collects rewards or penalties: winning move, point; losing move, punishment. Attempt after attempt, it refines its strategy. That’s how AlphaGo learned to beat human Go champions, and how robots get trained.

And generative?

Systems that create content — text, images, code — are the newest family: we covered them in the article on generative AI. Under the hood it’s still machine learning, at giant scale.

Practical example: your streaming service uses supervised learning to estimate how much you’ll like a film (millions of ratings serve as examples), unsupervised learning to discover “types” of viewers similar to you, and blends the results into your “recommended for you” row.

Where the method shines, and where it strains

Machine learning excels where rules are impossible to write: recognizing faces and voices, translating, estimating risk. But everything hinges on its raw material: a model is only as good as its data. Skewed or dirty examples produce skewed and dirty predictions, and a bias present in the data gets learned with the same diligence as any other pattern.

There’s also a more fundamental limit: the model detects correlations, it doesn’t grasp causes. When the world changes (new habits, new spammer tricks), the learned patterns age and predictions decay, until someone retrains it on fresh data.

How a machine actually “adjusts its parameters” while studying is the story of neural networks — the mathematical engine behind deep learning. And all of it, at bottom, remains what computing has always been: an algorithm grinding through well-defined steps — except this time, the steps were written by the training.