Why does the regularization term prevent overfitting in machine learning?

Is there any statistical interpretation for the regularization term?