VVZ API is not affiliated with ETH Zurich. Data might be outdated or incorrect. Please view the official ETHZ Vorlesungsverzeichnis for binding information.
Pattern matching beyond i.i.d. data
Last Updated: 2026-06-03 00:14:10
Abstract
The seminar will explore the theoretical and empirical properties of representations in generative AI.
Objective
The students gain an understanding of the challenges posed by distribution shifts and explore various techniques to effectively address them.They learn to understand a research paper in depth and to clearly communicate the main results to an audience.
Content
Recent advances in generative AI have enabled the learning of complex data distributions. However, deploying models beyond their training distribution remains challenging, particularly for tasks that require reasoning, prediction of unseen interventions, or operation in non-stationary environments. In this seminar, we will study the theoretical foundations of out-of-distribution generalization, focusing on causal approaches to address distribution shifts and on properties of learned representations that support adaptation to new environments. We will also examine recent empirical techniques for adapting models to novel distributions, including fine-tuning and few-shot learning and applications to domains such as music, biology, and language.
General Information
- Language
- English
- Levels
- MSC , WBZ
- Frequency
- Yearly recurring
Examination
- Type
- graded semester performance
Registration & Places
- Max Places
- 20
Course Components
| Type | Title | Time & Place | Hours |
|---|---|---|---|
| seminar |
Pattern matching beyond i.i.d. data
Does not take place this semester.
Blockseminar
|
No time listed | 2 h weekly |