Towards Worst-Case Guarantees with Scale-Aware Interpretability

Neural networks process information at multiple levels of detail simultaneously, similar to how natural data is organized hierarchically (from fine details to broad patterns). Current interpretation methods often fail to account for this multi-scale structure, potentially missing important aspects of how models make decisions.

The paper proposes using renormalization techniques from physics to create more rigorous interpretability tools that can formally guarantee bounds on how much influence discarded "noise" has on model outputs. This addresses a key gap in responsible AI: ensuring we can reliably explain what models are actually doing rather than relying on approximations.

Current interpretability methods lack formal robustness guarantees, meaning they may give misleading explanations of model behavior. The authors argue that borrowing mathematical frameworks from statistical physics could provide the theoretical foundation needed to build interpretation tools with provable reliability properties.

Scale-aware interpretability could improve AI safety by enabling researchers to verify that models aren't relying on spurious fine-grained patterns (like image artifacts or data anomalies) that might not generalize or could encode biases. By explicitly tracking how features combine across different resolutions, developers could better audit whether models are learning robust or fragile decision rules.