Rethinking Generalization: The Generalization Gap

The problem, from first principles

The generalization gap is how much better a model does on data it trained on than on data it has not seen. The paper's point is that this gap can be near 100 percent for random labels, while staying small for real labels.

From theory to code

Implement generalization_gap(train_acc, test_acc), returning the training accuracy minus the test accuracy.

Constraints

• Accuracies are between 0 and 1.

Hints