Deep Hierarchical Variational Autoencoder

The Nouveau variational autoencoder is a deep hierarchical VAE built for image generation that uses a stack of depth-wise separable convolutions and batch normalization to do the work.  Here is a link to the paper titled 'NVAE: A Deep Hierarchical Variational Autoencoder'. 

Yannic Kilcher runs us through an explanation of the paper below.

We recently had a HTC post on the followup architecture to this work you can check out here.

Also, make sure to check out Max Wellings lecture on unifying VAEs and Flow architectures here.

It's probably also time to dive back into David McAllester's presentation on VQ-VAEs.

And just to drive it all home, let's check out this 2020 lecture on variation autoencoders from Paul Hand.


Popular posts from this blog

Simulating the Universe with Machine Learning

CycleGAN: a GAN architecture for learning unpaired image to image transformations

Pix2Pix: a GAN architecture for image to image transformation