Course 4, lesson 32 of 100, Ages 10+
Features
The clues an AI looks at
Like I’m 5
To guess if a fruit is a banana, you look for clues: is it yellow, long and curved? Those clues are called features. AI uses features too.
The big idea
A feature is one measurable clue about an example. For a house, features might be its size, number of rooms and location. For an email, they might be the words used and the sender.
Choosing good features matters. Height is a great clue for guessing someone's shoe size, but their favourite colour isn't. Modern deep learning can discover useful features by itself from raw pixels or words.
Examples
- Fruit: Colour, length and curve help tell bananas from apples.
- House prices: Size, rooms and neighbourhood predict price better than the colour of the door.
- Deep learning: Image models learn their own features, like edges, fur and eyes.
How it works
- Pick clues that could help answer the question.
- Measure those clues for every example.
- Let the model learn which clues matter most.
Check your understanding
- Which is the most useful feature for predicting shoe size?
- Options: Height; Favourite colour; Birth month.
Answer: Height. Taller people tend to have bigger feet, so height is a useful clue. - What can deep learning do with features?
- Options: Discover useful features by itself from raw data; Only use features a person writes; Ignore all features.
Answer: Discover useful features by itself from raw data. Deep networks learn their own features from pixels, sounds or words.
Remember
Features are the clues a model uses. Good features make good predictions.
Talk about it
What features would you use to guess someone's favourite sport?
Go deeper
Hand-designing features (feature engineering) dominated classic machine learning. Representation learning lets neural networks learn features automatically, which is a key reason deep learning took off.