One of the most interesting aspects of deep learning has been finding ways to use neural networks without labelled data.
Autoencoders are one of the older ideas, but still highly applicable
Reducing the Dimensionality of Data with Neural Networks