Generalization
The whole point — doing well on data the model has never seen, not just the examples it trained on.
Think of it like
Learning to ride any bike, not just the exact one you practiced on.
Example
A well-generalized spam filter catches brand-new scam wording it never saw in training.
How it actually works
Generalization is the gap between training performance and real-world performance. Everything in ML — regularization, validation sets, more data — is ultimately in service of closing it. Overfitting and underfitting are the two ways it breaks.
For product teams
It’s the only performance number that matters — measure on data the model hasn’t seen.
For engineers
Expected performance on the true data distribution, estimated via held-out/test error.
Related
- Overfitting — Broken by overfitting.
- Underfitting — Also broken by underfitting.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome