The summer school is targeted toward PhD students working with data science broadly and for whom generative modelling potentially plays a part in their projects. The objective of the course is to introduce the students to the basics behind the most widely used deep generative models as well as expose the students to cutting-edge research in deep generative models. En studerende, der fuldt ud har opfyldt kursets mål, vil kunne:

• Comprehend, explain and apply the general concept of generative modelling, the likelihood function and maximum likelihood estimation.
• Comprehend, explain and apply the principles behind the most common deep generative models, including autoregressive models, flow-based models, deep latent variable models, energy-based models, generative adversarial networks and probabilistic circuits.
• Explain the assumptions and limitations of the most common deep generative models.
• Use a machine learning framework with automatic differentiation (e.g., PyTorch) to implement deep generative modes, including deep latent variable models, autoregressive models, and flow-based models.
• Document and disseminate software implementations of deep generative models.
• Give an overview of and compare the most common deep generative models.
• Match appropriate deep generative models to corresponding modelling problems.
• Discuss and disseminate how the most common deep generative models can be applied in their own research project.
• Relate the most common deep generative models to cutting-edge research in the field.

The course gives the students an introduction to the basics behind the most widely used generative modelling techniques and insights into the newest research on deep generative models and their applications. The course is accordingly divided into two parts.

The first part (Monday-Tuesday-Wednesday) consists of lectures and practical lab sessions. The topics covered within the first three days are 1) introduction to generative modelling, 2) autoregressive models, 3) flow-based models, 4) probabilistic PCA, 5) deep latent variable models, 6) energy-based models, 7) generative adversarial networks and 8) probabilistic circuits. The reading material for this first part is the course book and code examples listed as course literature. A detailed reading list is given on the course homepage, and this is mandatory preparation for the course. The course organisers conduct the lectures in the first part, and teaching assistants aid the lab sessions. This part is concluded by a mandatory poster session, where the participant share their (planned) research and discuss how deep generative models can be applied in the given context.

The second part (Thursday-Friday) will contain invited talks given by leading researchers in academia and industry. These talks will focus on both cutting-edge methodological developments in deep generative modelling and applications of deep generative modelling. Tomczak, Jakub M. Deep generative modeling. Springer, 2022.
https:/​/​doi.org/​10.1007/​978-3-030-93158-2

Online code examples:
https:/​/​github.com/​jmtomczak/​intro_dgm 19. marts, 2024