Generalization is a model’s ability to perform well on new inputs that were not present in its training data. A model learns a useful process rather than memorizing examples. section 1.27
A model that recognizes new images
An image categorization model should recognize a new dog or plane image, not only the labeled images it saw during training. New sentences and images are equivalent tests of whether it learned the task broadly. section 1.27
Generalization sits between two common failure modes:
- Overfitting occurs when a model memorizes particular training examples and performs poorly on unseen data. section 1.27
- Underfitting occurs when a model learns a process that is too simple to work effectively even on its training data. section 1.27
