Autoencoder

Data -> New representation -> original data

The idea is to convert the data into a latent representation (which is a representation in latent space), but makes the data more compact and interpretable.