The Same Machine, New Tricks
You probably know the saying: you cannot teach an old dog new tricks. But with learning machines, that is not true at all! The very same machine that learned to spot cats in photos could — with different training — learn to recognize handwriting, or translate languages, or predict the weather. Today we will find out why one machine can have so many different skills.
The Machine Is Like an Empty Notebook
Think of a learning machine as an empty notebook. The notebook does not care what you write in it. You could fill it with recipes, or stories, or math problems. The notebook itself does not change — only what you put inside it. A learning machine is similar. The machine itself is just a set of math and structure. What makes it useful is the examples you train it on. Train it on cat photos, and it becomes a cat finder. Train the same machine on recordings of music, and it becomes a music recognizer. Train it on descriptions of weather, and it becomes a weather guesser. The machine is the notebook. The training examples are what you write inside.
A learning machine does not come with a built-in job. Its job is decided by the examples you train it on. Change the examples, and you change what the machine can do.
Here is a story that shows this idea. Lena and her twin brother Leo each have the same type of tablet. They download the exact same learning app. Lena uses her app to practice recognizing animals — she shows it hundreds of animal photos labeled with names. Leo uses his app to recognize musical instruments — he shows it hundreds of instrument photos. After a month of training, Lena's app is great at naming animals. Leo's app is great at naming instruments. Same app, completely different skills — because they trained it on completely different examples. The app did not decide what to learn. Lena and Leo decided, by choosing the examples.
Match each training example set to the skill the machine would learn.
Terms
Definitions
Drag terms onto their definitions, or click a term then click a definition to match.
This is one of the most exciting things about learning machines. You do not need to build a brand new machine for every job. You can take the same kind of machine and train it fresh on new examples. Scientists have even discovered that sometimes you can take a machine that learned one thing and use that as a head start for learning something related. This is called transfer learning. It is like how being good at basketball helps you pick up volleyball faster, because some skills transfer.
When you learn to read in English and then start learning Spanish, you do not start from zero. You already know what letters are, what sentences do, and how to sound out words. Your reading skills transfer! Machines can do the same thing.
What decides what skill a learning machine will have?
Lena and Leo have the same app but trained it on different things. What will happen?
What Would You Train It On?
- Imagine you have a learning machine that starts as a blank slate — it knows nothing.
- Choose a skill you wish a machine could learn (for example: recognizing your pet, sorting your toys by type, or knowing what food you like).
- On paper, describe your training examples: what would you show the machine? What label would you give each example?
- Write down at least five specific example-and-label pairs.
- Share your training plan with someone: do they think the machine would learn your chosen skill from those examples?