When developing a deep learning model with a large number of features and uncertainty about their importance, which regularization technique would effectively drive the parameters of the least important features toward zero?