Efficient XAI Techniques: A Taxonomic Survey

Speed vs. Accuracy Trade-off is a Central Challenge: The paper identifies a critical tension in efficient XAI between making explanations faster to compute and keeping them faithful (accurate). Organizations deploying AI systems must understand that accelerating explanations may compromise how well they truly represent what the AI model is actually doing.

Two Main Approaches to Faster Explanations Exist: The survey categorizes efficient XAI into "non-amortized methods" (explaining one decision at a time) and "amortized methods" (learning patterns to explain multiple decisions quickly). This distinction matters for practitioners choosing which approach fits their real-world deployment needs.

Real-World Deployment Requires Careful Planning Across Three Phases: The paper highlights that efficiency challenges exist not just during explanation generation, but also during model training and deployment phases, plus in how explanations are actually used. This suggests responsible AI teams need end-to-end strategies, not just point solutions.

Current XAI Methods Are Too Slow for Many Real Applications: The paper's motivation—that "traditional XAI methods typically suffer from high computational complexity"—underscores why explainability hasn't been widely adopted in time-sensitive systems like autonomous vehicles or medical diagnostics, despite being crucial for responsible AI in these domains.

No One-Size-Fits-All Solution Exists Yet: The survey emphasizes challenges in "selection of different acceleration methods," indicating that organizations must carefully evaluate which efficient XAI technique matches their specific use case rather than assuming any acceleration method will work universally.