How Does AI Learn?

Learning by example

Think about how you learned to tell a cat from a dog. Nobody gave you a list of rules. You saw lots of cats and lots of dogs, and your brain learned the difference.

AI learns in a similar way. We show it lots of examples, and it finds the patterns.

If you are new to AI, start with Meet Artificial Intelligence.

Training data

The examples we give an AI are called training data.

To teach an AI to spot cats, we collect thousands of photos. We label each one: "cat" or "not a cat." 🐱

The AI studies all these labelled photos. It is training. The more good examples it sees, the better it gets.

Practice and feedback

AI does not learn perfectly the first time. It learns by practising.

Here is the loop:

1. The AI guesses an answer for an example.
2. We check: was it right or wrong?
3. If it was wrong, the AI makes a small adjustment.
4. It tries again, and again, thousands of times.

This is called getting feedback. Step by step, the AI's guesses get better. We call this machine learning, because the machine is learning from data.

Good data in, good AI out

Here is something important: AI only learns what we show it.

If we show it fair, correct examples, it learns well. If we show it wrong or unfair examples, it learns the wrong things. This problem is called bias.

For example, if an AI only ever saw photos of small dogs, it might not recognise a big dog. The data was missing something.

So people who build AI must choose their data carefully and fairly.

AI still does not "understand"

Even after all this learning, the AI does not understand a cat the way you do. It does not know a cat purrs or likes naps.

It only knows patterns in the data. That is why AI can still make silly mistakes, and why people should check its work.

Curious how the maths behind this works? When you are older, read What Is Machine Learning?.