Difference between flowtron and hierarchical generative GM-VAE by google

[artificertxj1]

Hi guys,
First, thanks for the great works.
My background is from computer vision and am not really familiar with sequential data deep learning and Tacotron details.
My major question when I read both papers, is what is the major difference between the two models?
Can I have some hints on that? Thanks.
Regards,
Justin Tian

[rafaelvalle]

Hey Justin,

Take a look at this document (https://arxiv.org/pdf/1908.09257.pdf) for a general comparison between normalizing flows and vaes.

For specific comparisons, take a look at our Flowtron paper(https://arxiv.org/pdf/2005.05957.pdf).

Generally speaking, normalizing flow's have an objective function that makes training simpler and more stable. In addition, by having a latent space with same dimensionality as the data, normalizing flows can store more information than vaes. This allows us to perform manipulations over time that are not possible in the conventional vae setup.