XAI2CSI: Interpreting CSI with eXplainable AI for Human Activity Recognition

Models fail to generalize because they learn environment-specific shortcuts rather than true activity patterns: The study found that deep learning models trained on Wi-Fi sensing data become overly dependent on context-specific signal patterns. When deployed in new environments or with different users, the models misinterpret underlying signal dynamics, revealing they've learned superficial correlations rather than robust activity recognition principles.

Explainability tools can diagnose why AI systems fail in real-world deployment: By using SAGE (a model explanation technique), researchers identified exactly which parts of the Wi-Fi signal data the model relied on for decisions. This diagnostic approach revealed generalization failures that wouldn't be apparent from accuracy metrics alone, demonstrating that explainability is essential for understanding when and why AI systems break down.

Transparency is critical for safety-sensitive sensing applications: Since Wi-Fi-based activity recognition could be used in homes or workplaces to monitor people, understanding model decisions becomes a responsibility issue. The paper's framework shows how to verify that systems are making decisions for the right reasons—not just achieving high accuracy through problematic shortcuts.

Current AI systems for wireless sensing need fundamental improvements before widespread deployment: The research reveals that existing deep learning approaches have "limited robustness to unseen conditions," suggesting these systems aren't ready for reliable real-world use without significant redesign. This finding cautions against deploying such technology without first addressing these fundamental generalization problems.