Generalization

Definition

Generalization is the ability of a model to perform usefully on new inputs, not only on the examples it trained on.

Simple form:

learn from training examples -> work on unseen examples

The important word is unseen.

Generalization means learned behavior transfers beyond the exact training cases.

---

Why Generalization Matters

A model can memorize training examples.

But memorization alone is not enough.

Real use requires the model to handle inputs that are different from the examples it saw.

Example:

training:
many labeled examples

real use:
new examples with unknown labels

The value of learning is measured by behavior on new cases.

---

Generalization vs Memorization

Memorization means the model stores or reproduces specific examples.

Generalization means the model learns reusable structure.

memorization:
works because this exact case was seen

generalization:
works because this case follows a learned pattern

Both can exist in the same model.

The problem is when memorization replaces useful pattern learning.

Generalization depends on reusable patterns, not only stored examples.

---

Connection To Representation

The previous chapter described embeddings as learned representations.

Generalization asks whether those representations are useful beyond the training set.

Example:

similar representation -> similar behavior on new inputs

If the representation captures useful structure, the model can apply that structure to new cases.

Representation supports generalization when it preserves useful patterns.

---

Generalization Is Not Guaranteed

A model can fit the training data and still fail on new data.

This can happen when it learns:

noise
shortcut patterns
accidental correlations
training-specific details

So training performance alone is not enough.

A model that learns the training set well has not necessarily learned the task well.

---

Final Summary

• Generalization means performing usefully on unseen inputs.
• It is different from memorizing exact examples.
• It depends on reusable learned structure.
• Embeddings and representations can support generalization.
• Training success does not guarantee generalization.