https://github.com/google-research/disentanglement_lib

Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations,ICML19

ICML2019 best paper

https://blog.acolyer.org/2019/07/17/challenging-common-assumptions-in-the-unsupervised-learning-of-disentangled-representations/

Important conclusion:

unsupervised models without conditioning in the latent space are in general unidentifiable.

Nonlinear ICA Using Auxiliary Variables,AISTAT19

In the following, we analyse our algorithm separately for the general case, and for conditionally exponential independent components of low order k. The fundamental result is that for sufficiently complex source distributions, the independent components are estimated up to component-wise nonlinear transformations; if the data comes from an exponential family of low order, there is an additional linear transformation that remains to be determined (by linear ICA, for example).