Explaning with trees: interpreting CNNs using hierarchies

Better explanations require faithful representation of how models actually work: The paper identifies a key problem with existing explanation methods like LIME—they can be intuitive to humans but may not accurately reflect what the neural network is actually doing. The proposed xAiTrees framework prioritizes "fidelity," meaning explanations genuinely represent the model's reasoning rather than just being easy to understand.

Hierarchical breakdowns help identify bias and unfair decisions: By providing "multiscale explanations" that break down decisions at different levels of detail, this approach can help researchers and auditors spot when AI systems are relying on biased patterns or irrelevant features—crucial for catching discriminatory outcomes before deployment.

One-size-fits-all explanations are insufficient: The framework supports both "human-centric" explanations (designed for people to understand) and "model-centric" explanations (showing what the model actually learned), recognizing that different stakeholders—regulators, engineers, affected communities—may need different types of clarity about AI decisions.

Current popular explanation methods have significant limitations: The paper notes that Integrated Gradients produces "noisy maps" that obscure understanding, highlighting that widely-used explainability tools may not be as reliable as practitioners assume, suggesting the need for continued innovation in this critical area.