In a Variational Autoencoders, z cannot be simply sampled from the output of the autoencoder directly since the network would not be differentiable. Instead, we have to sample from a normal distribution who's parameters are the output of the autoencoder, mu and stddev.

I So cannot understand how the network isn't differentiable (lets say we take z = mu + srd). So the reparameterization is not clear to me.

Any help would be much appreciated!!


1 Answer 1


Have you read Doersch's tutorial on VAE? In paragraph 1.1 the author shows main intuitions behind the latent model and in the following sections he goes deeper into the explanation of how a VAE works. I think it's a good starting point to approach to VAEs.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.