Computing Actual Causes for Neural Network Predictions under Structured Causal Inputs

Existing explanation methods can be misleading: Current popular techniques for explaining neural network predictions (like feature attribution) treat input features as independent, which produces incorrect explanations when real-world data has structured relationships between features—a common scenario that existing methods fail to account for.

Structured dependencies matter significantly: When the paper accounted for input dependencies using causal models, they found that 14.9% of explanations reported by traditional methods were actually spurious (false), demonstrating that ignoring these relationships leads to fundamentally flawed explanations.

Formal guarantees improve trustworthiness: By grounding explanations in Halpern-Pearl actual causes and Boolean Structural Causal Models, the approach provides mathematical completeness and minimality guarantees—meaning users can have formal assurance that all relevant causes have been found and no unnecessary ones are included.

The method scales to realistic complexity: The approach successfully handles large, complex scenarios with up to 28 variables and search spaces exceeding 2 trillion possibilities, suggesting it could be practically applied to real neural network explanations rather than remaining a theoretical exercise.

Responsible AI requires modeling real-world structure: This work highlights that trustworthy AI explanations cannot treat features in isolation; they must respect the actual dependencies and causal relationships present in real data to avoid systematically misleading stakeholders about why AI systems make their decisions.