In his work on Geomancer, David Pfau found that metric information was sufficient and necessary for disentangling but that learned embeddings were ...
In his work on Geomancer, David Pfau found that metric information was sufficient and necessary for disentangling but that learned embeddings were insufficient for doing so — he hoped that you could use another embedding or representation learning method and geomancer as a post-processing step to get disentangled subspaces. This didn’t work when he tried beta-VAE + geomancer and speculated it’s because nothing about VAEs forces them to preserve topology. What happens if you use something like a topology VAE instead?