Uploaded February 2021 | Updated November 2026, 2 hours ago
In the final video of this lecture we look at a more principled approach to building generative models: the variational autoencoder. We start with a simple maximum likelihood objective for our generator network, and with a series of approximations and simplifications, end up with an autoencoder that combines two loss terms and a sampling step to create a properly laid out latent space.
slides: mlvu.github.io/lectures/51.Deep%20Learning2.annotated.pdf
lecturer: Peter Bloem
In the final video of this lecture we look at a more principled approach to building generative models: the variational autoencoder. We start with a simple maximum likelihood objective for our generator network, and with a series of approximations and simplifications, end up with an autoencoder that combines two loss terms and a sampling step to create a properly laid out latent space.
slides: mlvu.github.io/lectures/51.Deep%20Learning2.annotated.pdf
lecturer: Peter Bloem










