Regularization
Also known as: regularize
Regularization is the umbrella for techniques that deliberately constrain a model so it generalizes better rather than memorizing.
Common forms:
- Dropout — random units off.
- Weight decay — penalize large weights.
- Data augmentation — vary the inputs.
- Early stopping — stop before full memorization.
The philosophy: perfect training-set fit is the *enemy*, not the goal. Regularization is how you trade a little training accuracy for real-world robustness.