Introduction: Painting Pictures with Neural Networks Using VAEs

In this chapter, we’ll discuss a class of generative models known as variational autoencoders (VAEs). These models are designed to make the generation of these complex, real-world images more tractable and tunable. They do this by using a number of clever simplifications to make it possible to sample over the complex probability distribution represented by real-world images in a scalable way.

We’ll explore the following topics to understand how VAEs work:

• How do neural networks create low-dimensional representations of data and some desirable properties of those representations?

• How do variational methods allow us to sample from complex data using these representations?

• How does using the reparameterization trick allow us to stabilize the variance of a neural network based on variational sampling, a VAE?

• How we can use inverse autoregressive flow (IAF) to tune the output of a VAE?

• How do we implement VAE/IAF in TensorFlow?