Stochastic Hi-Fi Just Uncovered AI's Hidden Reasoning — With a 411x Leap Over Scalar Limits
New method separates uniqueness, redundancy, and synergy in complex models like GPT-2
A new paper from researchers at the University of Bari and others exposes a fundamental flaw in common interpretability methods: scalar pairwise interaction scores (like Shapley interaction indices) conflate three distinct mechanisms—uniqueness (U), redundancy (R), and synergy (S). They prove this on a minimal 3-way XOR structural causal model, showing that faithful indices return zero per pair while projective indices spread third-order effects into conflated scalars.
To solve this, the team introduces Stochastic Hi-Fi, a post-hoc, retraining-free predictability decomposition that estimates per-feature U/R/S profiles through interventional masked inference with strict variance reduction (coupled diamond sampling). In tests on tabular SCMs, it recovers structure missed by scalar baselines by up to 411x. On GPT-2's indirect object identification (IOI) circuit, it separates redundant and synergistic attention heads. On NIH ChestX-ray14, Stochastic Hi-Fi matches GradCAM's localization (Pointing Game) and substantially improves Deletion AUC, indicating better faithfulness in interpreting model decisions.
- Proves that scalar pairwise scores like Shapley-Taylor conflate uniqueness, redundancy, and synergy, using a minimal 3-way XOR model
- Stochastic Hi-Fi achieves up to 411x larger interaction-magnitude recovery ratios compared to scalar baselines on tabular SCMs
- On GPT-2's IOI circuit, it separates redundant and synergistic attention heads; on NIH ChestX-ray14, it matches GradCAM and improves Deletion AUC
Why It Matters
Enables more faithful interpretability by decomposing feature interactions, crucial for debugging and trusting complex AI systems.