Visual Intelligence Online Seminar #79: Diffusion Model Meets XAI: Counterfactual Generation for Model Debugging
Presented by Nina Weng, PhD Candidate at the Technical University of Denmark
Counterfactual explanations provide a powerful way to understand model decisions by showing how inputs can be minimally altered to change predictions. Diffusion models offer a flexible and high-quality mechanism for generating such counterfactuals, enabling precise and realistic edits to image features. This makes them especially useful for model debugging tasks, such as detecting and quantifying shortcut reliance in medical imaging models.
In this talk, I will present our recent work on a fast, spatially constrained diffusion-based counterfactual framework (FastDiME), demonstrating its effectiveness in identifying model biases across multiple datasets.
It will be based on our work: Fast Diffusion-Based Counterfactuals for Shortcut Removal and Generation (ECCV 2024 oral, project page: https://fastdime.compute.dtu.dk/).