Solving the enigma: Enhancing faithfulness and comprehensibility in explanations of deep networks

Different explanation methods give conflicting results: The paper identifies a critical problem where established XAI methods often produce significantly different explanations for the same AI decision, creating uncertainty about which explanation to trust—a major barrier to responsible AI deployment in high-stakes fields like healthcare and medicine.

A new framework balances accuracy with understandability: The researchers developed an "explanation optimizer" that combines multiple XAI methods to create better explanations by simultaneously maximizing faithfulness (how accurately the explanation reflects what the AI actually did) and minimizing complexity (keeping explanations simple enough for humans to understand).

Substantial improvements in explanation quality: The framework achieved faithfulness scores 155% higher than the best existing methods in 3D medical imaging tasks and 63% higher in 2D object recognition, suggesting meaningful progress toward more reliable AI explanations.

Addresses a fundamental XAI limitation: By tackling the trade-off between accuracy and comprehensibility in explanations, this work helps resolve a key challenge in responsible AI—ensuring that explanations are both truthful about how the model works and understandable to non-technical stakeholders who need to trust AI decisions.

Tested on real-world critical applications: The validation on medical imaging (neuroscience) and object recognition demonstrates the framework's potential relevance for actual high-stakes domains where explainability is essential for responsible deployment.