The Query Channel: Information-Theoretic Limits of Masking-Based Explanations
Researchers have developed a new theoretical framework to understand the limitations of masking-based AI explanation methods like KernelSHAP and LIME. By modeling the explanation process as communication over a query channel, they identified information-theoretic limits on how accurately features can be identified. The study demonstrates that while information theory permits reliable explanations within certain query budgets, standard methods like Lasso and OLS may still fail, suggesting a gap between theoretical possibility and practical implementation. AI
IMPACT Establishes theoretical bounds for AI explainability, potentially guiding development of more reliable interpretation techniques.