X-REFINE: XAI-based RElevance input-Filtering and archItecture fiNe-tuning for channel Estimation

XAI can make critical wireless AI systems more trustworthy: The paper demonstrates that explainable AI techniques can help identify which parts of a neural network are actually important for making decisions in channel estimation—a critical function for 6G communications—making these "black-box" systems more transparent and reliable for real-world deployment.

Interpretability doesn't require sacrificing performance: X-REFINE maintains robust error rates (bit error rate performance) while simultaneously reducing computational complexity, showing that making AI systems more explainable and efficient are not competing goals but can be achieved together.

Internal model structure matters as much as input data: Unlike previous XAI approaches that only filter irrelevant input data, X-REFINE also optimizes the internal architecture of the neural network itself, suggesting that responsible AI requires examining both what data goes into systems and how those systems are internally organized.

Explainability enables practical deployment in critical infrastructure: By providing "high-resolution relevance scores" that identify the most reliable components of the model, X-REFINE addresses a key barrier to adopting AI in critical applications like wireless communications, where understanding why the system makes decisions is essential for safety and regulatory approval.