Autoencoder

Has an Encoder-Decoder-Architecture. We can use it to get rid of redundant information and compress the rest of the information in a nearly lossless way such that we can always reconstruct the original data from the compressed/auto-encoded data.

so here, we either learn matrices $A$ and $B$ or $A$ and $A^T$. This is in parts similar to PCA.

It can e.g. be used for Anomaly Detection.

Encoder

A module that compresses the input data into an encoded representation that is way smaller than the input data (Dimensionality Reduction).

Bottleneck

A module that contains the compressed data.

Decoder

A module that can decompress the compressed data and reconstructs the original data.

Linear Auto-Encoder

Types of Autoencoders

There are five popular types of autoencoders:

• Undercomplete Autoencoder
• Sparse Autoencoder
• Contractive Autoencoder
• Denoising Autoencoder
• Variational Autoencoder