Joint Explainability-Performance Optimization With Surrogate Models for AI-Driven Edge Services

Explainability and accuracy don't have to be enemies: The research demonstrates that AI systems can achieve over 99% explainability (measured by "Fidelity") while sacrificing less than 3% in predictive accuracy, suggesting that building interpretable AI doesn't require major performance trade-offs as previously assumed.

Surrogate models make complex AI understandable: By training simpler, human-readable models (like linear regression) alongside complex "black-box" AI systems, the approach enables stakeholders to understand why an AI made a specific decision—critical for applications like healthcare or financial services where decisions affect people's lives.

Joint training produces better explanations: Training the complex model and its simpler explanation model together simultaneously yields more trustworthy explanations than training them separately, suggesting that explainability should be built into AI systems from the start rather than added afterward.

Edge services need reliable AI: The focus on edge computing (AI running on local devices rather than distant servers) highlights a practical concern: as AI moves closer to users and makes real-time decisions with limited oversight, having trustworthy explanations becomes essential for safety and accountability.

This approach is practical for real-world deployment: The method balances competing demands (accuracy, interpretability, and computational efficiency) in a way that could work for actual AI systems serving users, rather than remaining a theoretical exercise.