What is machine learning?

Your phone suggests the next word as you type. Your email hides spam. A video app guesses what you will want to watch next. Nobody sat down and wrote a rule for each of these. The computer learned how to do them. That is machine learning.

Instructions versus learning

Most computer programs follow instructions that a person wrote, step by step. A calculator is like that. It adds numbers exactly the way it was told to.

But some jobs are too tricky to write instructions for. Try describing exactly how to tell a cat from a dog in a photo. Which pixels? What shapes? You would never finish, and there would always be a cat that breaks your rules.

Machine learning turns the problem around:

Instead of writing the rules, show the computer lots of examples (or let it practise), and let it work out the rules itself.

What does "learning" mean for a computer?

A learning program has a few simple parts:

1. Experience. Examples to look at, or a world to practise in.
2. A model. A machine with lots of adjustable numbers, a bit like dials. At the start, the dials are set at random, so it is no good at all.
3. A score. A way to tell whether an answer or a move was good or bad.
4. Practice. The computer tries, checks its score, and turns the dials a tiny bit to do better. Then it does that thousands or millions of times.

That is all learning is: nudging numbers so the score goes up. The computer does not understand cats or mazes the way you do. It has just found numbers that work.

Three ways to learn

1. From examples. Show the computer many pictures, each with a label ("apple", "banana"). It works out what apples tend to look like, and can then label new pictures. This is how photo apps and spam filters learn. It is called supervised learning, because the examples come with answers.

2. From rewards. The computer tries things and gets points when it does well. Over time it finds out which choices lead to more points. This is how programs learn to play games, and how some robots learn to walk. It is called reinforcement learning.

3. From patterns. The computer gets a pile of things with no labels and finds groups of similar ones by itself. Shops use this to spot kinds of customers. It is called unsupervised learning.

Try it: learn from rewards

In Train the Puppy, you meet the second way. A virtual puppy knows nothing about its maze. You set the rewards, and it learns the route by trial and error, using a real learning algorithm.

Machine learning is not magic

A learning program does exactly what its examples or rewards teach it, and nothing more:

• If the examples are poor or unfair, the answers will be poor or unfair. People say garbage in, garbage out.
• If the reward is badly chosen, the computer may find a sneaky way to score points without doing what you wanted. You will see this happen in the game.
• A learner can only be as good as the practice it gets.

That is why people who build AI spend so much time on the data and the rewards, not just on the code.