generalization

Definition

The property that matters most: performance on data the model never saw in training. Memorising the training set is trivial, so papers report the gap between training and held-out error, and any claim of progress is ultimately a claim about generalisation. When the test data comes from a shifted distribution, the term gets qualified — out-of-distribution or compositional generalisation — because ordinary held-out accuracy no longer covers that case.

Explainers using this term

• Statistical Learning Theory — Why Does Learning Generalize?Understanding Deep Learning Requires Rethinking Generalization