Deep Generative Modeling

This lecture introduces deep generative modeling as an unsupervised learning approach for estimating data distributions and generating new instances. It covers autoencoders, variational autoencoders, generative adversarial networks, latent spaces, image synthesis, and domain translation.

Syllabus

Intro
Supervised vs unsupervised learning Supervised Learning Unsupervised Learning
Generative modeling Goal Take as input training samples from some distribution and learn a model that represents that distribution
Why generative models? Debiasing
Why generative models? Outlier detection
What is a latent variable?
Autoencoders: background
Dimensionality of latent space reconstruction quality
Autoencoders for representation learning
Traditional autoencoders
VAEs: key difference with traditional autoencoder
VAE optimization
Intuition on regularization and the Normal prior
Reparametrizing the sampling layer
Why latent variable models? Debiasing
Generative Adversarial Networks (GANs)
Intuition behind GANS
Training GANs: loss function
GANs for image synthesis: latest results
Applications of paired translation
Paired translation: coloring from edges
Distribution transformations GANG