Autoencoders

Outline:

• Scratch
• Autoencoders

Scratch

• A Simple Autoencoder
• Convolutional Autoencoder

Autoencoders

• Type of network architecture are used to compress data, as well as image denoising.
  • compression and decompression functions are learned from data itself.
• General idea
  • pass an input data through an encoder, to make compressed representation
  • then pass the compressed representation through a decoder, to get back reconstructed data
  • encoder and decoder are both built with neural networks.
  • The whole network is trained by minimizing the difference between input and output data.
• Pros and cons
  • Pros
    • image denoising
    • dimensionality reduction
  • Cons
    • worse at compression. (jpeg, mp3 are better)
    • problems with generalizing to datasets.