Their method for training a shallow network requires that one first train a deep network.