The Polarised Regime of identifiable Variational Autoencoders

Lisa Bonheme, Marek Grzes

Keywords: variational autoencoders, VAEs, polarized regime, identifiability, iVAEs

TL;DR: An extension of the polarised regime to identifiable VAEs

Abstract: The polarised regime—the capacity of variational autoencoders (VAEs) to discard superfluous latent variables—is well-studied in the context of “classical” VAEs with a standard Gaussian prior. In this paper, we extend these results to the case of identifiable VAEs (iVAEs).