What does it mean for a machine learning model to generalize?

Generalization means that a machine learning model performs well on new, unseen data that is similar to the data it was trained on.

A model is useful when it learns underlying patterns rather than simply memorizing the training examples.

Simple example

Suppose a model is trained to predict house prices using historical data.

If it performs well only on the houses it saw during training, it has not learned enough to be useful.

If it can also make accurate predictions for new houses it has never seen before, the model is generalizing well.

Training performance is not enough

A model can achieve very low error on its training data and still perform poorly on new data.

This often happens when the model learns details or noise that are specific to the training set instead of learning broader patterns.

That problem is called overfitting and is covered separately.

What good generalization looks like

A model that generalizes well should:

• perform well on unseen validation or test data
• learn patterns that apply beyond individual training examples
• avoid depending too heavily on noise or accidental details in the training set

Generalization does not mean a model will work equally well on completely different kinds of data. It usually assumes that future data is reasonably similar to the data used during development.

In short, a machine learning model generalizes when what it learned from the training data also works well on new, unseen examples.