generalization

Generalization is a model's ability to do well on data it has never seen — not just on the examples it studied. It is the entire point of machine learning. A child who has met a few dogs can recognize a breed they have never encountered; that leap from specific examples to a working rule is generalization. A model that can only repeat the answers to its practice questions has learned nothing useful, the way memorizing one exam paper teaches you nothing for a different exam.

We measure it honestly by hiding some data from the model during training and then testing on that held-out set. Strong performance on the training examples but weak performance on the held-out ones means the model has overfit — it memorized quirks and noise instead of the real pattern. The reverse problem, doing poorly even on the training data, is underfitting: the model is too simple to capture the pattern at all. Good generalization threads between the two.

Crucially, generalization is never guaranteed and never unconditional. A model generalizes only to data that resembles what it was trained on. Train a model on sunny-day photos and it may fail in the rain; train it on one hospital's patients and it may stumble at another. This breakdown when the world shifts — distribution shift — is one of the most common and underappreciated ways real systems fail. Generalization is a hopeful bet, supported by careful testing, not a promise.