“Explainability of Vision-Language Models”
In this talk, I will introduce the methodology for explaining vision-language models and discuss the key challenges associated with this research domain. The presentation will highlight our recent scientific results concerning the FIxLIP algorithm published at NeurIPS’25, which enables a better understanding of multimodal encoder predictions. Following this, I will provide a brief demonstration of how to apply game theory to estimate feature attributions of predictive models using the {shapiq} Python package.
Bio
Hubert Baniecki is a PhD student at the University of Warsaw and a research assistant at the Centre for Credible AI (CCAI). He researches machine learning interpretability and explainable AI within the group led by Prof. Przemysław Biecek. During his PhD, Hubert gained experience as a visiting researcher at LMU Munich. He is a scholar of the Foundation for Polish Science and the Ministry of Science and Higher Education, and has co-authored scientific results published at leading AI conferences (NeurIPS, ICML, ICLR).