L17.4 Variational Autoencoder Loss Function

Slides: sebastianraschka.com/pdf/lect...
-------
This video is part of my Introduction of Deep Learning course.
Next video: • L17.5 A Variational Au...
The complete playlist: • Intro to Deep Learning...
A handy overview page with links to the materials: sebastianraschka.com/blog/202...
-------
If you want to be notified about future videos, please consider subscribing to my channel: / sebastianraschka

[@desucam7717]

The best lecture I've heard in terms of VAE. Very much thanks!!!

[@SebastianRaschka]

Thanks so much for the kind words! Feels really good to hear this :)

[@satadrudas3675]

The explaination and reasoning on the reconstruction loss was simple and concise. Thanks

[@SebastianRaschka]

Glad to hear!

[@mustafabuyuk6425]

Thank you for a very clear explanation.

[@SebastianRaschka]

Glad to hear it was useful!

[@hophouse8426]

Thank you for good lecture! but I have one question. Reconstruction term in ELBO is expectation value about z~q(z|x). But when you use MSE loss, it does not seem to be considered. Is there any reason??

[@miguelduqueb7065]

Thank you for the explanation. In minute 7:54, I think the formula for the squared error loss (the L-2 norm) should not contain the squared root.

[@jiangshaowen1149]

Great Lecture! I have a naive question, what's the difference between q and p in Encoder and Decoder? Do they refer to conditional probability? Thanks

[@klindatv]

Is correct to say that with binary cross entropy the decoder will generate slightly different samples instead of using the MSE? So the decoder output is different than the original input?

[@SebastianRaschka]

yes, the loss function is a hyperparameter and the results will be slightly different.

[@maulberto3]

Hi, why the KL divergence between the Z and a Normal(0,1) distribution?

[@maulberto3]

I think it's a way to tell the NN to produce nice images, therefore normal, and not produce weird ones as the normal has not big tails. So I guess that if I choose say the t-distribution then I will let the NN produce more weird images than the Normal architecture....

[@SebastianRaschka]

For a normal distribution, we know how to sample from it. And it's easy to work with mathematically (via the log-var trick). But you could also consider other distributions of course.