zettelkasten

Auto Encoders

Last updated: 1/9/2025

Questions

Notes

Autoencoders are types of neural networks that comrpess the data into a latent space that can be used to reconstruct the previous data.

Types

Denoising autoencoder

It is an autoencoder that takes in a noisy image and then outputs a "cleaned" version of the image.

Variational autoencoders

These are types of autoencoders that are better at generating data because the latent space is smaller and more clustered.

They have a higher reconstruction loss compared to regular autoencoders, but they are more fault tolerant (their latent space can be represented with less precision).

You can use VAE's to increase the size of your dataset.

VqVAEs

Applications

It can be used to generate more examples in supervised learning (i.e. make more training data)

See Also

  1. [[sigaida]]