What is machine learning?

In plain English

Machine learning is a way of building software that learns from examples rather than following instructions a programmer typed out. You give the system lots of data — past emails marked spam or not spam, say — and a learning procedure gradually adjusts the system until its guesses match the known answers reasonably well. Once trained, it can make guesses about new cases it has never seen.

There are a few common flavours. Supervised learning uses labelled examples (photos tagged “cat” or “dog”). Unsupervised learning looks for structure in unlabelled data, such as grouping customers into similar clusters. Reinforcement learning improves by trial and error against a score, the way a program learns to play a game.

Why it matters

Machine learning is what made AI practical. Many useful tasks are impossible to write rules for — nobody can list every rule that distinguishes a cat from a dog in a photo — but they are easy to demonstrate with examples. Virtually every AI product you hear about today, from fraud detection to chatbots, is machine learning underneath.

An example

A bank wants to spot fraudulent card payments. Rather than writing rules by hand, it feeds the system millions of past transactions, each marked genuine or fraudulent. The system learns which combinations of amount, location, time and merchant type tend to signal fraud, and starts flagging suspicious new payments.

What to watch out for

A machine learning system is only as good as its examples. If the historical data reflects past bias or covers only certain kinds of customers, the system will reproduce those gaps — and do so with an air of mathematical authority. It also finds correlations, not causes: it may learn that a harmless detail happens to accompany fraud, and then fail when conditions change.