Why do neural networks need a separate validation set?

Why

A model predicting drug response may appear accurate when repeated measurements from the same patient are split across training and validation. That is leakage: the model can recognize patient-specific patterns without generalizing to new patients. Split by the real unit of deployment, which could be patient, donor, site or study, and perform normalization and feature selection within each training fold.

Use validation results for learning rate, architecture and early stopping, then freeze the selection strategy before final testing. Reusing a test set repeatedly makes it a de facto validation set. When possible, evaluate a time- or site-separated cohort to test transfer.

State whether the benchmark asks about new specimens from known people or truly unseen individuals, because these are distinct generalization targets.