Overall Course Objectives
To provide knowledge of current research topics in generative modeling, including handling issues of identifiability and non-trivial data such as graphs.
- Explain in detail how deep generative models work.
- Operationalise and implement deep generative models.
- Compare and distinguish the modeling choices in different deep generative models.
- Reason about which aspects of a statistical model yields identifiable outcomes.
- Operationalize differential geometric representations in latent variable models.
- Estimate identifiable distributions from observational data expressed in a learned representation.
- Operationalize and implement graph neural networks.
- Reason about the foundations of graph neural networks.
The course consists of 3 modules which each consist of 2-3 weeks of lecturing and two weeks of advanced project work. The topics are deep generative models, geometric representations, and graph neural networks. The focus is on the theoretical foundations and mathematical model components.
02450/02456/02476/02477/02405, This course is an advanced course in machine learning. Students are expected to have passed most of the machine learning courses offered by DTU Compute before attending this course. At a minimum, students are expected to have passed 02450: Introduction to Machine Learning and Data Mining and 02456: Deep Learning. Students will benefit most from the course if they have also passed 02476: Machine learning operations, 02477: Bayesian machine learning, and 02405 Probability Theory. The course can be followed in parallel with 02477.
Lectures with exercises, and periods of project work.
This course is an advanced course in machine learning and part of the focus area Machine Learning and Signal Processing of the Master of Mathematical Modelling and Computing program.