Why is analytically solved closed Kl divergence not used at least for the gaussian prior ?

ThirstyScholar / bayes-by-backprop

" Weight Uncertainty in Neural Networks"

https://arxiv.org/abs/1505.05424

45 stars 9 forks source link

Open Pran97 opened 5 years ago

Alaya-in-Matrix commented 5 years ago

The author of the paper claimed that the usage of monte carlo ELBO approximation enables more expressive prior.