Generalization
Also known as: generalize
Generalization is the real test: can the model do well on things it has never seen? That’s what separates genuine learning from memorization.
A model that only works on its training examples is useless in the wild. Generalization is why we:
- Hold out a test set the model never trained on.
- Measure performance on that set, not the training set.
- Talk about “robustness” and “out-of-distribution” performance.
The north star: every metric that matters in the end is a generalization metric — how the model behaves on the new stuff you actually care about.