XAI for Black-Box Models Using Logic-Based Formal Methods and Combinatorial Optimization – FORM-XAI

The FORM-XAI project addresses the growing need for interpretability in the field of artificial intelligence (AI), particularly in critical sectors such as healthcare, finance, and transportation. Complex models, such as deep neural networks, while effective, raise concerns about their opacity, thereby limiting user trust. The objective of this project is to develop explainable AI (XAI) approaches that provide rigorous, coherent explanations tailored to user needs while maintaining high performance.

By relying on formal methods and combinatorial optimization techniques, FORM-XAI aims to make black-box models more transparent. Unlike current agnostic methods, which often lack coherence and do not consider users' cognitive limitations, the proposed approach focuses on generating explanations with mathematical guarantees. However, to overcome the high computational cost associated with these explanations, the project proposes efficient approximations that balance rigor and speed.

The adopted methodology will be interdisciplinary, integrating techniques from propositional logic, statistical sampling, and submodular optimization, thereby ensuring verifiable and interpretable explanations. The project also includes a validation phase based on real-world case studies, targeting areas where transparency is paramount.

The scientific value of FORM-XAI lies in the combination of formal approaches and optimization to enhance the transparency of deep learning models. This project addresses contemporary regulatory issues related to AI and the increasing concerns around trust by developing solutions that ensure the accountability of AI systems.