Bengio et al (2013) note that techniques for greedy, layer wise, unsupervised, pre-training were a significant technical breakthrough, enabling deep learning systems to learn a hierarchy of features one level at a time, using unsupervised feature studies to learn a new transformation at each level