Calendar

Trial Lecture for appointment as Associate Professor - Tuwe Löfström-Cavallin

You are invited to attend Tuwe Löfström-Cavallin's trial lecture for appointment as Associate Professor 

TITEL
When Is a Prediction Enough?
A model that cannot express uncertainty is not ready to support decisions.

ABSTRACT
A model can perform well on average while leaving us without enough information to act on a particular prediction. What supports its answer? How uncertain is it? And how much confidence should we place in the explanation?

This lecture explores these questions through two recurring themes in my research: explanations and uncertainty. Beginning with my first paper on explaining machine-learning models in 2004, I connect approaches that reveal broad patterns in model behaviour with explanations of individual predictions. Along the way, I discuss why differences between models can improve predictions, how uncertainty can be quantified and calibrated, and what statistical guarantees tell us about predictive reliability.

The main focus is Calibrated Explanations, which brings local explanations together with calibrated predictions and uncertainty information. Through visual examples, I show how uncertainty can change our interpretation of an explanation, how the question being asked affects what is explained, and how alternative model outcomes can be explored. The aim is to clarify what explanations and uncertainty contribute to assessing whether a prediction provides enough support for a decision, and where their limits lie.

Welcome!

Link to Zoom-Webinar: https://ju-se.zoom.us/j/61195359798