Respect the model: Fine-grained and Robust Explanation with Sharing Ratio Decomposition

Existing AI explanations may not accurately reflect how models actually work: The paper challenges the truthfulness of current explanation methods, showing they have "deviated from faithfully representing the model," which means users may be getting misleading information about why AI systems make their decisions.

AI explanations can be manipulated through adversarial attacks: Current explanation methods are "susceptible to adversarial attacks," meaning bad actors could potentially trick these systems into providing false explanations—a significant concern for trustworthy AI deployment in high-stakes applications.

The proposed SRD method provides more robust and honest explanations: By adopting a "vector perspective" rather than focusing only on individual neurons, SRD "sincerely reflects the model's inference process" and shows "significantly enhanced robustness," offering a more reliable way to understand AI decision-making.

Inactive neurons matter as much as active ones: The paper introduces an important observation that both active and inactive neurons contribute to model decisions, requiring a redefinition of what "relevance" means—this suggests previous explanations may have missed half the story about why models behave as they do.

Fine-grained explanations are now possible at multiple levels: SRD enables "recursive decomposition" and provides detailed explanations at any layer of the model, allowing stakeholders to understand AI decisions at different levels of detail depending on their needs.