AI-Fun with ELLIS Seminar | Yusuf Sale: Online Selective Conformal Inference: Errors and Solutions
| Dates: | 25 March 2025 |
|---|
| Times: | 14:00 - 15:00 |
|---|
| What is it: | Seminar |
|---|
| Organiser: | Faculty of Science and Engineering |
|---|
| Who is it for: | University staff, External researchers, Alumni, Current University students |
|---|
| Speaker: | Yusuf Sale |
|---|
|
Title:
Online Selective Conformal Inference: Errors and Solutions
Abstract:
Conformal inference provides a principled way to construct prediction intervals with guaranteed coverage. However, in online selective conformal inference, where predictions are made only when a selection rule is met, things become trickier — selections can break the exchangeability of the data, leading to incorrect coverage guarantees.
In this talk, I will critically examine existing calibration selection strategies used to restore exchangeability and highlight some fundamental errors in their theoretical claims regarding selection-conditional coverage and false-coverage statement rate (FCR) control. To address these shortcomings, I will introduce new calibration selection strategies that maintain exchangeability, ensuring both valid selection-conditional coverage and FCR control. Beyond the theory, I will discuss practical trade-offs between different approaches and present experimental results that illustrate their effectiveness.
Bio:
Yusuf Sale is a final-year PhD student at the Chair of Artificial Intelligence and Machine Learning (AIML) at LMU Munich, supervised by Prof. Eyke Hüllermeier. He is a member of the Konrad Zuse School of Excellence in Reliable AI (relAI), an initiative supported by the Federal Ministry of Education and Research.
His research focuses on uncertainty quantification in machine learning, with a particular emphasis on the representation and quantification of different types of uncertainty (viz. aleatoric, and epistemic uncertainty). His work aims to develop reliable and trustworthy AI and ML systems that effectively account for inherent uncertainties in their predictions, thereby enhancing decision-making across a wide range of (high-stake) applications. Moreover, he is interested in distribution-free uncertainty quantification, exploring both its theoretical foundations and practical applications.
As ever, in-person attendance is encouraged but if you are unable to attend in person, you can register (link on this page) to receive a link to join online.
Speaker
Yusuf Sale
Organisation: LMU Munich
Travel and Contact Information
Find event
6.210
University Place
Manchester
