Learning From Examples

How did you learn what a dog looks like? Nobody handed you a rulebook that said: four legs, fur, tail, barks. You just saw dogs — at the park, in books, in cartoons — over and over. Your brain soaked up all those examples and built a picture of what a dog is. Machines can learn the same way! Today we will find out how showing a machine lots of examples teaches it to recognize things in the world.

Examples Are a Machine's Teachers

When scientists want to teach a machine to recognize cats, they do not write a list of cat rules. Instead they collect thousands — sometimes millions — of photos of cats. Each photo has a label attached to it. The label says: this is a cat. The machine looks at all those labeled photos. It searches for patterns. What do cats usually have? Pointy ears. Whiskers. A certain shape of face. The machine does not see those features the way you do. It sees numbers that represent colors and shapes. But it finds patterns in those numbers. After studying enough examples, the machine can look at a brand-new photo it has never seen before and say: yes, that is a cat, or no, that is not.

Labels matter just as much as examples. Imagine you are learning what an apple is, but every time someone shows you an apple they call it a banana by mistake. You would get very confused! The same thing happens with machines. If a photo of a cat is labeled dog by mistake, the machine learns something wrong. That is why scientists are very careful to make sure their examples are labeled correctly. A good example has two things: the picture (or data), and the correct label. Together they are called labeled data. Labeled data is the food that a learning machine eats.

Let us look at a real example of this process. A team of scientists wanted a machine to tell spam emails from real emails. They collected ten thousand old emails. They labeled each one: spam or not spam. The machine studied all ten thousand labeled emails. It found patterns: spam emails often had words like win and free and urgent. Real emails from friends had different words and styles. After studying the examples, the machine could look at a new email and make a good guess: spam or not spam. The scientists did not write a single spam rule. The machine found the patterns itself, from the examples.