Rajesh Ranganath

Spring 2018

Day/Time: Thursdays 5:10-7:00pm

Location: CIWW 312

Generative models define procedures that produce samples of data. They can be used to learn representations, to handle exploration/exploitation tradeoffs, and to make use of the large amounts of unlabeled data. Deep generative models use ideas from deep learning to build generative models and algorithms for learning them. This course will focus on some of the recent advances in deep generative models. Students will embark on a semester-long research project around deep generative models.

**Introduction and Logistics****Generative Models and Variational Inference**

**Autoregressive Models and Invertible Transformations**

**Adversarial Learning**

**Evaluation of Generative Models****What’s up with exchangeability? (Guest: Victor Veitch)****Information Theoretic Views**- Elements of Information Theory, 2nd Edition, Chapters 1 + 2
- The Information-Autoencoding Family: A Lagrangian Perspective on Latent Variable Generative Modeling

**Geometry of Deep Generative Models****Cross Domain Sample Matching****Structured Models and Inference****Model-Based Reinforcement Learning****Sequence Models (Guest: Emily Denton)****Mixture**