Stacked autoencoders are DNNs that are typically used for data compression. Their particular hourglass structure clearly shows the first part of the process, where the input data is compressed, up to the so-called bottleneck, from which the decompression starts.
The output is then an approximation of the input. These networks are not supervised in the pretraining (compression) phase, and the fine-tuning (decompression) phase is supervised:
Figure 11: Stack autoencoder architecture