E.T. : Empirical Theory in Representation Learning

About

This workhop is about promoting scientific theory building in deep representation learning; see the synopsis for more information. We have related international keynotes, a call for papers, which are pre-registered and peer-reviewed and after acceptance will be published in official procedings.

• Date/time: Sep 8, 2026 at 13:15 CEST (Europe/Stockholm time)
• Location: Quality View Hotel - Stroget D, Malmö, Sweden
• All posters will be in the Malmo Massan Exhibit Hall.
• Link to virtual workshop site

 

Workshop schedule

13:15	Opening	Jan van Gemert	Head Computer Vision lab, TU Delft	Motivation
13:30	Keynote:	Wieland Brendel	Principal Investigator (PI) Max Planck Institute for Intelligent Systems	45 min
14:15	Orals		see below	45 min
15:00	posters		coffee break	1h
16:00	Keynote:	Margret Keuper	Full Professor for Machine Learning at Mannheim University	What makes models robust?
16:45	Keynote:	Steven Scholte	Neuroscientist at University of Amsterdam	Small theories of large models: predict, control, extend

 

Oral presentations:

• “The FID Lottery: Quantifying Hidden Randomness in Generative Model Evaluation” by Nicolas Dufour, Alexei A Efros, Patrick Perez

• “How Image Classifiers Accumulate Class Evidence with Depth, and How to Supervise It” by Han Wang, Hilde Kuehne

• “When Is a Task Vector Enough? An Empirical Theory of Implicit Multimodal ICL” by Jiaqian Li

• “How neural network architecture shapes the reliance on local features” by Aurélien Boland and Hannah Pinson.

• “Probing Feed-Forward 3D Reconstruction through Controlled Representation Intervention” by Matti Marino Schlenker, Fran Zezelj, Philipp M. H. Bernhardt, Bernhard Schölkopf, Andreas Geiger, Polina Karpikova, Gege Gao

• “Where Does Generative Difficulty Reside? An Empirical Study of Target Representations” by Marcel Plocher, Bernhard Schölkopf, Andreas Geiger, Gege Gao

---