When rules run out: why computers need to learn

You have learned a lot about how computers work. They take in bits, follow clear instructions, and do billions of tiny steps a second. If you can write down the steps, a computer can do the job.

But what if you cannot write down the steps?

Jobs that rules handle well

Some jobs are perfect for rules:

• Add up a shopping bill.
• Sort a list of names alphabetically.
• Find the shortest route on a map.
• Play noughts and crosses perfectly.

For each of these, a clear list of steps always works. That is what normal programming is brilliant at.

Jobs where rules fall apart

Now try this: write rules to decide whether a photo shows a cat.

• "A cat has pointy ears." But so do foxes, and some cats have folded ears.
• "A cat has whiskers." You cannot see them in many photos.
• "A cat has four legs." Not if it is curled up asleep.
• "A cat is furry." So is a dog. And some cats are hairless.

Every rule you write has exceptions, and every exception needs another rule. You could write thousands and still be wrong.

It is the same for many jobs that people find easy:

• Reading handwriting. Everyone writes a 7 differently.
• Spotting spam email. A rule like "block emails that say free money" is beaten when spammers write "fr3e m0ney".
• Understanding speech. Voices, accents and background noise are all different.
• Driving a car. The world has too many surprises.

We know how to do these things, but we cannot explain our own steps. Try telling someone how you recognise your friend's face. You cannot, you just do.

A different idea: learn from examples

Instead of writing the rules, what if the computer found them itself?

1. Collect lots of examples as a table of numbers (remember: everything can be numbers).
2. Tell the computer what counts as a good answer.
3. Let the computer practise, adjusting its own settings to do better.

Then the computer finds its own rules, hidden in thousands of numbers, that no person had to write. This is machine learning.

Learning is not always better

It is not magic, and it is not always the right tool:

• A calculator is perfect with rules. Nobody wants a calculator that "learns" 2 + 2 and sometimes gets it wrong.
• Learning needs lots of examples, and good ones.
• It can make mistakes and be hard to explain.
• It does what its data and rewards teach it. That is why you saw the puppy find loopholes.

Good engineers pick the right tool for each job. Often they combine the two.

You are ready

You now know the building blocks: bits, numbers, gates, processors, algorithms, programs and data. That is enough to understand what machine learning is really doing. In the next lesson, you will meet it.