Deep Generative Modeling

Explore deep generative modeling in this lecture from MIT's Introduction to Deep Learning course. Learn about supervised vs unsupervised learning, outlier detection, and latent variable models. Dive into autoencoders, Variational Autoencoders (VAEs), and their optimization techniques. Discover Generative Adversarial Networks (GANs), including progressive growing of GANs and style-based generators. Examine domain transformation with CycleGAN and gain a comprehensive understanding of deep generative modeling techniques and applications.

Intro
Which face is fake?
Supervised vs unsupervised learning
Why generative models? Outlier detection
Latent variable models
What is a latent variable?
Autoencoders: background
Dimensionality of latent space → reconstruction quality
Autoencoders for representation learning
VAEs: key difference with traditional autoencoder
VAE optimization
Priors on the latent distribution
VAEs computation graph
Reparametrizing the sampling layer
VAEs: Latent perturbation
VAE summary
Generative Adversarial Networks (GANs)
Intuition behind GANS
Progressive growing of GANS (NVIDIA)
Style-based generator: results
Style-based transfer: results
CycleGAN: domain transformation
Deep Generative Modeling Summary