Machine learning

What do the photos show?

Name what each picture shows as quickly as possible. 3…2…1…GO! This is incredibly hard for a computer. But as you’ve already discovered, it can learn to recognise things – using machine learning!

Machine learning is an important part of AI. There are three types of machine learning.

1 Supervised learning

Step 1

You give the computer a lot of data and say what it is each time. For example, a lot of photos of chihuahuas and muffins, which you label “muffin” or “chihuahua” each time. Giving data to the computer to learn from is what we call training AI.

The computer then looks for patterns in the two groups of photos. Because they are somewhat similar, you will have to give it a lot of training. Give it a try here.

Step 2

Finished training? Now show the computer a new photo. Based on the patterns it has learnt, it will predict which answer is more likely: chihuahua or muffin.

95% chihuahua
The computer is 95% sure that it’s a chihuahua.

• In this case, the computer learns from labelled data.

2 Unsupervised learning

Okay, had enough of chihuahuas and muffins? There’s a lot more you can do apart from teaching a computer the difference between chihuahuas and muffins.

For example, you can also give a computer loads of data without saying what it is. It will then look for patterns and create categories itself. For example, loads of pancakes and waffles.

The pancakes are collected together and the waffles are collected together. Can you identify any other categories? (Check the background of the images.)

• In this case, the computer learns from unlabelled data.

3 Reinforcement learning

There is one more important method of learning. You might be familiar with it from computer games. With a game, you’re usually not that good to start with. But as you go along, you learn what helps you and which moves aren’t so good. You’re rewarded when you’re doing well – LEVEL UP – and penalised when you’re doing badly – GAME OVER. In this way, you get better all the time. And yes, you can make a computer learn things in the same way. For example, it can learn to play a game or ... find the way through a maze. Check out the screens below.

• In this case, the computer learns from data that it obtains by doing things well or badly. To put it another way, you can learn by giving things a try.

Deep learning

Digital brain

Within machine learning, there’s also an aspect called “deep learning”.

Deep learning is a complex part of machine learning that works specifically with neural networks. And what are they inspired by? The way our brains work.

This is what a neural network looks like

Lots and lots of data

A neural network is usually trained with enormous volumes of data: a huge amount of text, images or videos, which AI companies get from the Internet. Getting large quantities of data from web pages is called “scraping”.

Neural networks tend to summarise the key characteristics of all that data in a “latent space”.

Black box

Once our neural network is fully trained, you can feed it data and it will produce a result. The problem is that we don’t always know exactly what it has learnt. Often, we don’t know why it comes up with a particular result. What happens between data and result is therefore a “black box”.

If we don’t understand why an AI model produces a particular result, is it okay to use it? Suppose an AI model is trained to decide which study programmes are suitable for you, and it decides that the programme you wanted to take isn’t right for you – but nobody knows why. Would you go with its decision?