search

julian-8897 / hyperbolic_vae

Variational Autoencoder with non-euclidean (hyperbolic) latent space
12 stars 2 forks source link

What is the conclusion here? #1

Open xiaoiker opened 5 months ago

xiaoiker commented 5 months ago

Any insights about the the value of the HVAE? Can it be better than Euclidean ones?

julian-8897 commented 5 months ago

Hey there, thanks for the message, I read your survey paper a couple of year ago, it was very insightful! Although i left this repo quite some time ago, what I've found was that the representations learned by the HVAE within the latent space seems to be much more robust and demonstrated much better reconstruction quality compared to the Euclidean one , even for lower number of latent dimensions. I'm not entirely convinced how useful this might be, but it'll be interesting to see if the representations learned from the HVAE might be useful for some vision downstream tasks. It'll be good if this project can be revived somehow

xiaoiker commented 4 months ago

Thanks for the reply. I am always interested in extend generative models into hyperbolic space. That is the reason I am asking.

julian-8897 commented 4 months ago

Nice, thanks! Would be interesting to extend hyperbolic spaces beyond the VAE, do you have any thoughts on this? I'm also interested in seeing how we can quantify the learnt representations in hyperbolic space