Traditional software follows rules written by programmers. Machine learning flips that: instead of writing the rules, we show the computer many examples and let it find the patterns itself.
A simple example
Imagine sorting email into “spam” and “not spam”. Writing rules by hand (“if the message contains free money…”) quickly becomes impossible. A machine learning system instead looks at thousands of emails already labelled by people and learns which features tend to signal spam. When a new email arrives, it applies what it learned.
The three main types
- Supervised learning — learning from labelled examples. Photos labelled “cat” or “dog” → a model that recognizes cats and dogs.
- Unsupervised learning — finding structure in unlabelled data. Grouping customers with similar shopping habits.
- Reinforcement learning — learning by trial and error with rewards. A program that learns to play a game by being rewarded for winning.
What “training” means
During training, the model makes predictions, compares them with the right answers and adjusts its internal numbers (called parameters) to make fewer mistakes. Repeat this millions of times and the model becomes good at the task — as long as the examples represent the real world well.
Where you already use it
- Recommendations on streaming and shopping sites
- Spam filters and fraud alerts
- Photo search on your phone
- Voice assistants and automatic subtitles
- Navigation apps predicting traffic
Its limits
A model is only as good as its data. If the examples are biased or incomplete, the model will be too. And a model trained for one task usually cannot do another without new training.
What to read next
Machine learning is the foundation of today’s generative AI, which learns from huge amounts of text, images and audio in order to create new content.