Conceptual
Login

Training and Held-Out Data Splits for Evaluating Generalization

Why labeled data is partitioned into a portion used for fitting and a portion withheld from fitting, and why performance on the withheld portion is the measure of whether a model generalizes rather than memorizes.