Short Definition
Generalization is a model's ability to perform well on new data, not just the data used during training.
Intuition
A useful model learns the pattern behind the examples, not only the examples themselves.
Technical Definition
Generalization measures expected performance on unseen samples drawn from the target distribution, usually estimated with validation or test data.
Example
A speech model generalizes if it handles new speakers and accents that were not in the training examples.
Common Misunderstandings
Generalization cannot be proven by training metrics alone.
A model can generalize in one deployment setting and fail after distribution shift.