Variational Autoencoders: The Nuts and Bolts

Variational Autoencoders (VAEs) are a powerful generative model with many applications ranging from drug discovery to image de-noising. They can be used both to generate synthetic data as well as to discover latent codes of a data distribution. First introduced in 2013 by Kingma and Welling (“Autoencoding Variational Bayes”), VAEs are thought of as the other major type of generative model, an alternative to GANs.

In the first half of this talk, I will colloquially convey the general idea behind VAEs, and in the second half I will delve into their mathematical underpinnings. We’ll conclude by dissecting some VAE sample code in Keras.