/iwae_tutorial

Tutorial and discussion on Importance Weighted Autoencoder (IWAE) / Variational Autoencoder (VAE) implementation on MNIST using Tensorflow 2.0

Primary LanguageJupyter NotebookMIT LicenseMIT

Study on Importance Weighted Variational Autoencoders (IWAE)

The jupyter notebook included in this repository covers an implementation of Importance Weighted Autoencoders (IWAE) as detailed in https://arxiv.org/abs/1509.00519.

The implementation is done in Tensorflow 2.1.0 and tested on the binarized MNIST dataset.

The following theoretical and experimental explorations are covered in this implementation:

  • Theory behind IWAE
  • Choice of initialization strategy
  • Representation learning capabilities of IWAE
  • Effect of tighter IWAE bounds
  • Large-scale IWAE bound performance