Fast Fourier Correlation is a Highly Efficient and Accurate Feature Attribution Algorithm from the Perspective of Control Theory and Game Theory

Neural networks learn efficiently through frequency patterns: The research demonstrates that neural networks rely heavily on low-frequency features, and remarkably, only 8% of these Fourier features are needed to maintain accurate predictions for 80% of images in Vision Transformers—suggesting AI models are far more efficient than previously understood.

Current explanation methods may be incomplete: Traditional methods for explaining AI decisions focus on spatial features (like pixel locations), but this work shows that analyzing features through a frequency/signal perspective provides more accurate and complete explanations of what neural networks actually learned.

Better tools exist for AI transparency: Game-theoretic evaluation metrics are more appropriate for analyzing frequency-based features than spatial ones, meaning researchers should adopt different evaluation frameworks depending on the type of feature being examined—this matters for building trustworthy AI systems.

Fourier-based features are more interpretable and distinctive: Features identified through frequency analysis show stronger patterns within the same category and clearer differences between categories compared to traditional spatial methods, making it easier to understand and verify that AI systems are learning meaningful distinctions rather than spurious correlations.

Implications for efficient and explainable AI: These findings suggest that future AI systems could be made more efficient (requiring fewer features) and more interpretable by explicitly incorporating frequency-domain analysis, potentially leading to AI that is both faster and easier for humans to understand and audit.