NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·May 21·8 min read
GRALIS: A Unified Canonical Framework for Linear Attribution Methods via Riesz Representation
TL;DR
Popular neural network explanation methods (GradCAM, SHAP, LIME, Integrated Gradients) lack a unified mathematical foundation, making them theoretically incomparable. GRALIS establishes a canonical representation theory proving every additive, linear, continuous attribution functional (a mathematical function assigning importance scores to inputs) admits a unique representation via Riesz representation theorem.
✦ Why It Matters
Engineers can now rigorously compare and validate neural network explanation methods using a shared mathematical foundation instead of ad-hoc approaches.
Key Takeaways
- Use GRALIS framework to formally compare different XAI attribution methods mathematically.
- Understand that all major attribution methods fit one unified canonical representation.
- Apply Riesz representation theory to design new attribution techniques with guaranteed properties.
- Verify attribution method behavior through formal theory rather than empirical testing alone.
Full Summary
Explainable AI (XAI) attribution methods—techniques that identify which input features most influence a neural network's predictions—have proliferated without shared theoretical grounding. GradCAM, SHAP, LIME, and Integrated Gradients each operate on distinct mathematical principles, preventing rigorous comparison.
GRALIS (Gradient-Riesz Averaged Locally-Integrated Shapley) establishes a representation theory proving that every additive (contributions sum to total prediction), linear (proportional to input changes), and continuous (small input changes yield small attribution changes) attribution functional on L² function spaces admits a unique canonical representation. The framework leverages Riesz representation theorem from functional analysis to map attribution methods to a common mathematical space.
This unification enables engineers to understand relationships between methods, predict when methods agree or diverge, and potentially design new attribution techniques with guaranteed properties. The work provides formal guarantees about attribution method behavior rather than empirical comparisons alone.