![[2411.08875] Causal Explanations for Image Classifiers](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2411.08875] Causal Explanations for Image Classifiers
Summary
This paper presents a novel approach to generating causal explanations for image classifiers, introducing a black-box algorithm grounded in the theory of actual causality, and demonstrating its efficiency compared to existing tools.
Why It Matters
Understanding how image classifiers make decisions is crucial for transparency and trust in AI systems. This research provides a principled method for generating explanations, which can enhance the interpretability of AI models and support their deployment in sensitive applications.
Key Takeaways
- Introduces a black-box approach for causal explanations in image classifiers.
- Grounded in formal definitions of causality, enhancing theoretical rigor.
- Demonstrates that the proposed tool, ReX, outperforms existing methods in efficiency and explanation quality.
- Proves termination and discusses complexity of the algorithm.
- Provides experimental results validating the effectiveness of the approach.
Computer Science > Artificial Intelligence arXiv:2411.08875 (cs) [Submitted on 13 Nov 2024 (v1), last revised 19 Feb 2026 (this version, v3)] Title:Causal Explanations for Image Classifiers Authors:Hana Chockler, David A. Kelly, Daniel Kroening, Youcheng Sun View a PDF of the paper titled Causal Explanations for Image Classifiers, by Hana Chockler and 3 other authors View PDF HTML (experimental) Abstract:Existing algorithms for explaining the output of image classifiers use different definitions of explanations and a variety of techniques to find them. However, none of the existing tools use a principled approach based on formal definitions of cause and explanation. In this paper we present a novel black-box approach to computing explanations grounded in the theory of actual causality. We prove relevant theoretical results and present an algorithm for computing approximate explanations based on these definitions. We prove termination of our algorithm and discuss its complexity and the amount of approximation compared to the precise definition. We implemented the framework in a tool ReX and we present experimental results and a comparison with state-of-the-art tools. We demonstrate that ReX is the most efficient black-box tool and produces the smallest explanations, in addition to outperforming other black-box tools on standard quality measures. Comments: Subjects: Artificial Intelligence (cs.AI) Cite as: arXiv:2411.08875 [cs.AI] (or arXiv:2411.08875v3 [cs.AI] for this ve...
