Course 5, lesson 48 of 100, Ages 11+
Synthetic data
Made-up examples that help AI learn
Like I’m 5
Sometimes it's hard to collect real examples, like photos of rare accidents. So people use computers to make pretend examples. That's synthetic data.
The big idea
Synthetic data is created by simulations or other AI models instead of being collected from the real world. A driving simulator can create thousands of rainy night scenes without anyone being in danger.
It's useful for rare events and for protecting privacy, because no real person's records are used. But if the made-up data isn't realistic enough, the AI can learn habits that fail in the real world, so it's mixed with and tested on real data.
Examples
- Driving simulators: Virtual cities create rare, risky scenarios safely.
- Privacy: Fake patient records with realistic patterns protect real patients.
- Training chatbots: Some models learn from carefully checked AI-written examples.
How it works
- Build a simulation or generator that mimics the real world.
- Create many examples, especially rare ones.
- Test the AI on real data to check it still works.
Check your understanding
- Why use synthetic data for self-driving cars?
- Options: It can create rare, dangerous situations safely; It's always more accurate than reality; Cars prefer it.
Answer: It can create rare, dangerous situations safely. Simulation covers risky cases without real danger. - What's the main risk of synthetic data?
- Options: It may not match the real world well enough; It's illegal; It's always too small.
Answer: It may not match the real world well enough. Unrealistic data can teach habits that fail in reality, so real tests are vital.
Remember
Synthetic data fills gaps and protects privacy, but must be checked against real data.
Talk about it
What rare situation would you want a simulator to create for a robot?
Go deeper
Approaches include physics simulators, domain randomisation and generative models. Training models repeatedly on their own outputs without care can degrade quality, sometimes called model collapse.