XAI for Self-supervised Clustering of Wireless Spectrum Activity

Self-supervised learning reduces labeling burden while maintaining performance: The paper addresses a key practical challenge in AI deployment—supervised deep learning models require massive amounts of labeled data. Self-supervised approaches can achieve comparable performance with significantly less labeled training data, making AI systems more feasible to implement in resource-constrained settings.

Explainability methods can be combined to demystify complex AI systems: Rather than treating deep learning as an irreducible "black box," the researchers demonstrate that multiple explanation techniques (Guided Backpropagation for feature identification and decision trees for clustering logic) can be layered together to make AI reasoning transparent and auditable.

Visualization of AI decision-making enables human oversight: By connecting clustering decisions back to specific input features through data-specific visualizations, the methodology allows domain experts to verify whether the AI system is reasoning based on legitimate patterns or spurious correlations—a critical requirement for responsible deployment in sensitive applications like wireless communications.

Explainability is essential for self-supervised systems: While self-supervised learning solves the data scarcity problem, the paper highlights that explainability challenges persist. This suggests that responsible AI development requires addressing multiple concerns simultaneously—not just performance or efficiency, but also interpretability and trustworthiness.