Generalization

A student can memorise every practice answer and still struggle when the real exam changes the questions.

A network faces a similar test. Performing well on relevant examples it did not train on is called .

Suppose you have 10,000 labelled cat-and-dog photos. Set 2,000 aside before training. Train the network on the other 8,000.

The held-back photos must stay separate from the training examples. They should also resemble the kinds of photos the network will meet later.

After training, test the network on those unseen photos. Strong performance is evidence that its learned pattern works beyond the examples it saw during training.

Good training results show that the network fits its practice examples. Good results on new, representative examples show that the pattern travels.

When success on the training examples does not carry to new ones, the gap has its own name.