Learning from Less: Improving and Understanding Model Selection in Penalized Machine Learning Problems