Autoencoders are neural networks with same input and output. They are latent-factor models


  • Includes a bottleneck layer: with dimension smaller than input .
  • First layers “encode” the input into bottleneck.
  • Last layers “decode” the bottleneck into a (hopefully valid) input


  • Superresolution
  • Noise removal
  • Compression

Relationship to principal component analysis (PCA):

  • With squared error and linear network (no non-linear ), equivalent to PCA.
  • Size of bottleneck layer gives number of latent factors in PCA.