X-RACE: XAI-assisted Recurrent neural network Attribution for Channel Estimation

Making AI systems more trustworthy through transparency: The paper addresses a critical gap in responsible AI by developing methods to understand how LSTM neural networks make decisions in wireless communications. Rather than treating these systems as "black boxes," X-RACE provides explanations of which data inputs and internal components actually matter for predictions, making the technology more auditable and trustworthy.

Efficiency gains without sacrificing performance: X-RACE achieves a 44.1% reduction in computational complexity while maintaining or improving accuracy (measured by Bit Error Rate). This demonstrates that making AI systems more interpretable doesn't require accepting worse performance—a key concern for real-world deployment of responsible AI systems.

New metrics for understanding AI behavior over time: The paper introduces three novel metrics (Saturation Time, Importance Drift, and Relevance Contrast) specifically designed to track how LSTM networks learn and remember information. These temporal metrics represent progress toward better monitoring and understanding of AI system behavior, which is essential for responsible deployment.

Practical, scalable approach to explainability: Unlike traditional XAI methods that require expensive, repetitive computational processes, X-RACE uses a "one-shot dual-optimization strategy" that is faster and more practical. This addresses a real barrier to responsible AI adoption—making explainability techniques feasible for actual deployment rather than just research settings.