Deep Generative Models

[Important!] You should register for the seminar through the central proseminar assignment system.

Description: Generative models have been adopted in many AI/ML applications, such as computer vision, natural language processing, and scientific discovery. Recent advances in parametrizing these models using deep neural networks and training them using stochastic optimization methods have enabled scalable modeling of complex, high-dimensional data. In this proseminar, we will look into foundational frameworks and research frontiers of deep generative models. We will cover topics such as variational autoencoders, generative adversarial nets, diffusion models, flow-based models, and autoregressive models.

Each student will be assigned a topic and will present two papers related to the topic. Each presentation will be followed by a group discussion about the presented content, where you will receive constructive feedback on how to improve after the first presentation. Your final grade will be largely decided by your performance on the second presentation. At the end of the semester, each student will hand in a short two-page summary of the presented topic.

Time & Location: 16:15 - 17:45 weekly on Mondays; Room 0.07, CISPA Main Building

• Oct 27: Kick-off – Introduction to the proseminar and overview of topics

Requirements: The required language for the presentations is English. There are no formal requirements for this proseminar, but having taken graduate-level machine learning or optimization courses will be helpful. You are particularly welcome to the course if you are interested in doing research on deep generative models.