Pretty sure all this says is to minimize KL-Divergence instead of log-likelihood (for the encoder), or your latent variables are garbage. Judging by many examples I've seen of VAEs in ML, this is not news.<p>This post [0] does a good job showing the difference between latent variables for log-likelihood and KL losses.<p>[0]<a href="https://towardsdatascience.com/intuitively-understanding-variational-autoencoders-1bfe67eb5daf" rel="nofollow">https://towardsdatascience.com/intuitively-understanding-var...</a>
Getting a resource exhaustion error from Namecheap, here's a cache:<p><a href="http://web.archive.org/web/20190129144610/http://paulrubenstein.co.uk/variational-autoencoders-are-not-autoencoders/" rel="nofollow">http://web.archive.org/web/20190129144610/http://paulrubenst...</a>
What about variational homoencoders tho? See <a href="https://arxiv.org/abs/1807.08919" rel="nofollow">https://arxiv.org/abs/1807.08919</a>