New Framework Uses Formal Causality Theory to Explain Image Classifier Decisions

Image classification systems have become widespread, but understanding why they make particular decisions remains challenging. Most existing explanation tools rely on ad-hoc methods rather than a unified theoretical foundation.

A Principled Approach to Explanations#

A new paper proposes grounding image classifier interpretability in the formal theory of actual causality. Rather than borrowing informal explanation concepts, the researchers ground their work in rigorous causal definitions—offering the first black-box method to bridge the gap between causality theory and practical model explanations.

The approach comes with theoretical guarantees: the authors prove key results about their method and provide an algorithm for computing approximate explanations. They also establish when the algorithm terminates and analyze how the approximations compare to exact causal definitions.

ReX: A More Efficient Alternative#

The team implemented their framework in a tool called ReX and benchmarked it against state-of-the-art explanation tools. The results show ReX outperforms existing black-box approaches on multiple fronts:

• Generates the smallest explanations overall
• Offers superior efficiency compared to other black-box methods
• Achieves stronger performance on standard quality benchmarks

The work, submitted to arXiv in November 2024 and updated through July 2026, demonstrates that applying formal causal reasoning can produce explanations that are not only theoretically sound but also practically superior to heuristic alternatives.