Workshop: Generative Neural Networks with Applications to Imaging
Event details
| Date | 23.02.2024 |
| Hour | 13:00 › 15:00 |
| Speaker | Prof. Dr. Ullrich Köthe, Heidelberg University |
| Location | |
| Category | Conferences - Seminars |
| Event Language | English |
<< THE EVENT IS SOLD OUT >>
The workshop is jointly organised by the EPFL Center for Imaging and the EPFL AI Center.
Abstract:
Deep generative models have emerged as a powerful paradigm for hard problems in computer vision, natural language processing and the sciences. "Generative" here refers to two capabilities: (1) generate synthetic data that are (ideally) indistinguishable from real data, and (2) calculate the probability density of any given data point. These models thus acquire a high degree of understanding of the phenomenon under study, which can be exploited for novel solutions to high-dimensional probability modeling, Bayesian inference, and interpretability, to name just a few.
The first part of the talk will report about leading approaches to generative modeling (normalizing flows, rectangular flows, and diffusion models), introduce their theoretical and algorithmic foundations, and describe their most successful realizations by neural networks. It will be shown that a categorization in terms of change-of-variables formulas associated with these methods allows for a systematic understanding of many differences and similarities.
The second part will be devoted to applications of generative models as a tool for other machine learning tasks, e.g. generative image classification and the solution of inverse problems. Most notably, generative models allow for a fully Bayesian treatment of these tasks and therefore lead to trustworthy uncertainty quantification of the solutions. Examples from a wide range of applications, from medical imaging to psychology, will illustrate the power of the resulting framework.
Bio:
Ullrich Köthe received the Diploma degree in physics from the University of Rostock, Rostock, Germany, in 1991, and the Ph.D. degree in computer science from the University of Hamburg, Hamburg, Germany, in 2000.,He is currently an Adjunct Professor of computer science with the Interdisciplinary Center for Scientific Computing, Heidelberg University, Heidelberg, Germany. His research focuses on the connection between machine learning and the sciences from a methodological perspective and an application perspective and, in particular, on the interpretability of machine learning results.
Registration required
The workshop is jointly organised by the EPFL Center for Imaging and the EPFL AI Center.
Abstract:
Deep generative models have emerged as a powerful paradigm for hard problems in computer vision, natural language processing and the sciences. "Generative" here refers to two capabilities: (1) generate synthetic data that are (ideally) indistinguishable from real data, and (2) calculate the probability density of any given data point. These models thus acquire a high degree of understanding of the phenomenon under study, which can be exploited for novel solutions to high-dimensional probability modeling, Bayesian inference, and interpretability, to name just a few.
The first part of the talk will report about leading approaches to generative modeling (normalizing flows, rectangular flows, and diffusion models), introduce their theoretical and algorithmic foundations, and describe their most successful realizations by neural networks. It will be shown that a categorization in terms of change-of-variables formulas associated with these methods allows for a systematic understanding of many differences and similarities.
The second part will be devoted to applications of generative models as a tool for other machine learning tasks, e.g. generative image classification and the solution of inverse problems. Most notably, generative models allow for a fully Bayesian treatment of these tasks and therefore lead to trustworthy uncertainty quantification of the solutions. Examples from a wide range of applications, from medical imaging to psychology, will illustrate the power of the resulting framework.
Bio:
Ullrich Köthe received the Diploma degree in physics from the University of Rostock, Rostock, Germany, in 1991, and the Ph.D. degree in computer science from the University of Hamburg, Hamburg, Germany, in 2000.,He is currently an Adjunct Professor of computer science with the Interdisciplinary Center for Scientific Computing, Heidelberg University, Heidelberg, Germany. His research focuses on the connection between machine learning and the sciences from a methodological perspective and an application perspective and, in particular, on the interpretability of machine learning results.
Registration required
Links
Practical information
- General public
- Free
Organizer
- EPFL Center for Imaging, EPFL AI Center