Deep Dive Workshop
Interpretability in the lab and in the wild
Organized by:
Friday 9 October 10.00
Lead organizer: Lenka Tětková, DTU Compute
Interpretability has become a standard part of the data scientist’s toolkit. Explainability libraries are easy to install, but there is a gap between what they promise in papers and what they deliver on real models under real-time pressure. Explanations sometimes mislead, contradict, or reveal bugs you wish you had not found.
This session brings together practitioners and researchers to share what interpretability looks like in practice: e.g., a model debugged through an explanation revealing a hidden shortcut; a deployment where stakeholders disagreed on what makes a good explanation; a method that worked on benchmarks but broke in production. We want the successes and the failures, the tips that are not in any tutorial, and the open problems no one has solved.
Invited and lightning talks from industry and academia, a panel discussion, and roundtables will spark discussion and networking.
10:00 – 10:05
Welcome
10:05 – 10:45 Invited talks
Real world interpretability in hearing health care
Niels Henrik Pontoppidan, Principal Scientist, Eriksholm Research Centre
Hierarchical Reasoning Models, Deconstructed
Lukas Galke Poech, Associate Professor, University of Southern Denmark
10:45 – 11:10 Lightning talks, part 1
Simple sequence patterns support complex evolutionary representations in Evo 2
Rachael Marie De Vries (KU)
When the model disagrees: visualizing what ECG networks learn
Ahcène Boubekki (KU)
Learnable Time-Frequency Masks for Explaining Time-Series Classifiers
Theresa Dahl Frehr (DTU)
Under the Influence: Data Attribution for Transparent and Fair AI
Johannes Gabriel Sindlinger (ITU)
Characterizing Learning in Deep Neural Networks using Tractable Algorithmic Complexity Analysis
Pedram Bakhtiarifard (KU)
11:10 – 11:30
Structured discussions/round tables
11:30 – 12:30
Lunch break
12:30 – 12:50 Invited talk
Visual information flow in vision-language models
Isra Salazar, PhD student, KU
12:50 – 13:10
Panel discussion with all the speakers
13:10 – 13:35 Lightning talks, part 2
When Information-Elicitation Strategies Become Hard to Tell Apart
Stavroula Asteriou (KU)
Hierarchical Datasets for Foundation-Model Embedding Interpretability
Marco Parola (AAU)
What Cosine Similarity of Label Representations Can and Cannot Tell us
Beatrix M. G. Nielsen (ITU)
Model Sensitivity to Unexpected Regions of Medical Images
Théo Sourget (ITU)
Archetypal Analysis for Interpretable Characterization of Medications based on Adverse Drug Reaction Profiles
Espen Jimenez Solem (Bispebjerg & Frederiksberg Hospital)
13:35 – 13:55
Structured discussions/round tables
13:55 – 14:00
Wrap-up
Intermediate: For attendees who have basic understanding or some experience with the subject but are not yet advanced.