This project features some exploration to get a fully parameterized Gaussianization scheme. There is a big normalizing flows community with many different algorithms for density estimation and sampling. There is also a relatively small community using Gaussianization and density destructors for other applications including information theory measures. This is an attempt to bridge the two communities together.
This project was inspired by:
- RBIG - original algorithm
- Gaussianization Flows - the fully parameterized method.
- nflows - the research normalizing flows library.
Here is an example where we show the original data and the data generated by a trained Gaussianization Flow model.
Original Data | Generated Samples |
---|---|
![]() |
![]() |
The same Gaussianization Flow model with the initial Latent space versus the latent space after trained.
Original Latent Space | Trained Latent Space |
---|---|
![]() |
![]() |
Here is an example where we show the original data and the data generated by a trained Gaussianization Flow model.
Original Data | Generated Samples |
---|---|
![]() |
![]() |
...and the probabilities of the dataset.
Original Data |
---|
![]() |
The same Gaussianization Flow model with the initial Latent space versus the latent space after trained.
Original Latent Space | Trained Latent Space |
---|---|
![]() |
![]() |
This work was supported by the European Research Council (ERC) Synergy Grant “Understanding and Modelling the Earth System with Machine Learning (USMILE)” under Grant Agreement No 855187.