Undercomplete Autoencoder

A Unsupervised Learning Autoencoder that encodes unlabeled data and tries to predict (decode) to that same data.

Mostly used to generate the Latent Space which can be seen as a space that represents the data in lower dimensions (Dimensionality Reduction).

Principle Component Analysis can only recognize linear relationships and is thus not powerful enough to work with non-linear data.

Uses the Reconstruction Loss.