Title
Towards machine learning models that know what they do not know
Abstract
Machine learning models are increasingly deployed in high-stakes settings as decision support systems. A central challenge in such a setting is to design models that not only make accurate predictions, but also recognize when their own uncertainty should trigger alternative actions, e.g., abstaining from providing a prediction, deferring to a better predictor, or informing the decision maker about their own uncertainty.
In this talk, I will discuss recent work on making uncertainty actionable. I will first consider when a model should abstain, focusing on structured prediction settings. I will then discuss how uncertainty can guide human-AI interaction, from deferring decisions to requesting feedback or interventions when they are most useful. Finally, I will examine how better uncertainty quantification, especially through local calibration, can support these decisions. Together, these directions highlight both progress and open challenges in building models that know what they do not know.
Bio
Andrea Pugnana is an Assistant Professor in the Department of Information Engineering and Computer Science at the University of Trento.
Speaker is presenting in person in Tromsø.
Practical
The Tuesdays seminars series are devoted to various topics relevant for the Integreat´s research focus. Presenters from the Integreat community and beyond have 40 minutes to present, followed by a group discussion.
Seminars are open for attendance for everybody.
For those unable to attend in person: https://uio.zoom.us/j/62535744013
Point of contacts:
- Organiser and moderator: Veronica Lachi, UiT The Arctic University of Norway
- Integreat Administrative Leader Maria Dikova, University of Oslo