Machine Learning (ML)

Normally, humans tell a computer exactly what to do: step by step, in the form of rules. With machine learning, it works the other way around. You show the program a great many examples and let it find the rules itself. Someone who wants photos of cats and dogs sorted by animal type doesn’t describe what a cat looks like. Instead, they supply tens of thousands of images along with the correct answer, and the program adjusts itself until it usually classifies correctly. Machine learning, abbreviated ML, is the term for this approach.

Why hand-written rules aren’t enough

For many tasks, no one can write down the rules, even though every human masters them. You recognize your friend’s handwriting instantly, but you couldn’t explain exactly how. It was precisely these kinds of problems that computers failed to solve for decades. ML solves them because it replaces description with examples.

There is also an economic reason. Rules go out of date, whereas data keeps arriving constantly. An ML system for credit card fraud keeps learning from every new fraud case. A rigid rule set, by contrast, would have to be maintained by programmers, and the fraudsters would be faster.

That’s why ML is now the core of almost every AI product. When the news talks about “artificial intelligence,” it practically always means machine learning. AI is the umbrella term, ML is the method that actually worked.

From training data to a finished model

It starts with a model: a mathematical function with many adjustable numbers, the parameters. At the beginning these numbers are random, so the model is just guessing. Now it receives an example, makes a prediction, and that prediction is compared with the correct answer. The gap between the two is called the error.

This error is calculated backwards: each parameter is nudged a small step in the direction that reduces the error. This repeats millions of times. You can picture it like descending a mountain in fog — you only see the ground right beneath you and keep heading downhill. This phase is called training, and it can take anywhere from hours to weeks.

The test afterward is crucial. The finished model is checked against examples it has never seen during training. Otherwise overfitting occurs: the model has memorized the training data but fails on anything new. It’s like a student who memorizes old exams without understanding the material.

ML in apps, market reports and job postings

You encounter ML constantly in everyday life, usually invisibly. Your inbox’s spam filter is an ML model. So are the recommendations on streaming platforms, facial recognition when unlocking your phone, and automatic translation in the browser. Chatbots like ChatGPT are ML systems too, just very large ones.

In business news, certain terms around ML come up again and again. “Deep learning” refers to ML with especially many-layered models, so-called neural networks. “GPU” stands for specialized computing chips that speed up training — demand for them has made the chipmaker Nvidia one of the most valuable companies in the world.

A common misconception is that ML systems are objective because they work with data. The opposite can be true. A model absorbs the biases of its training data. If hiring software was trained on decades of decisions in which mostly men were hired, it will continue to favor men. Such cases are documented and are one of the reasons for AI regulation in the EU.