New Metric and Method for Quantifiable XAI Without Ground-Truth
A recent framework suggests a measurable approach for assessing Explainable AI (XAI) techniques in the absence of ground truth, utilizing continuous input perturbation. This metric formally addresses the sufficiency and necessity of the information attributed to model decisions, offering a closer alignment with human intuition compared to current metrics. Additionally, the authors present an innovative XAI technique that refines a model by employing a differentiable approximation of the metric as guidance, leading to the development of an adapter module that enables black-box models to provide causal explanations.
Key facts
- arXiv:2605.18681v1
- Announce Type: new
- Abstract: Explainable AI (XAI) techniques are increasingly important for validation and responsible use of deep learning models.
- Proposes a framework for quantifiable metric for XAI methods based on continuous input perturbation.
- Metric considers sufficiency and necessity of attributed information to model's decision-making.
- Metric aligns better with human intuitions of explanation quality than existing metrics.
- Novel XAI method fine-tunes a model using differentiable approximation of the metric as supervision signal.
- Result is an adapter module that can be trained on top of any black-box model to output causal explanations.
Entities
Institutions
- arXiv