Business Analytics Seminar: Cees Diks (UvA)
"Addressing underspecification in machine learning"
Event details from: Business Analytics Seminar: Cees Diks (UvA)
- Date
- 1 March 2022
- Time
- 12:00 – 13:00
Introduction:*
In a recent paper D’Amour et al. (2020) show that many failures of real-world ML can be attributed to underspecification, which is defined as the presence of many parameter combinations that achieve an equivalent generalization performance on an i.i.d. testset. They provide ample empirical evidence that underspecification can cause a significant drop in performance. Motivated by this, we recently initiated a project to explore how to address underspecification in real-life ML. The seminar will be informal, presenting to the audience which possible ways to detect and address underspecification we have explored so far, what the preliminary conclusions are, and which future approaches might be promising. The main aim of the presentation is to stimulate discussion and exchange of ideas on this topic.
Co-authored with Floris Holstege (EUR), Noud van Giersbergen (UvA) and Bram Wouters (UvA)
Contact:
This seminar will be organised via Zoom. If you are interested in joining this seminar, please send an email to the secretariat of Amsterdam Business School at secbs-abs@uva.nl.
