A technique used to prevent overfitting in neural networks by adding a penalty term to the error function.