Hierarchical Variational Autoencoder for Visual Counterfactuals

Counterfactuals can help explain AI decisions: The paper demonstrates that Hierarchical Conditional VAEs can generate "counterfactuals"—modified versions of images showing what would change if a specific feature were altered—providing a concrete way to understand why an AI classifier made a particular decision.

More complex AI models require new explanation techniques: While simpler conditional VAEs work for explainability, real-world applications need more sophisticated "Hierarchical" models. The paper shows that standard conditioning methods fail with these complex models, requiring new technical approaches (relaxing posterior effects) to generate reliable explanations.

Visual auditing of classifiers becomes possible: The introduced VAEX model enables "visual auditing" of image classifiers, allowing practitioners to systematically test how a classifier responds to specific feature changes—a practical tool for identifying potential biases or unexpected behaviors before deployment.

Theoretical soundness matters for trustworthy explanations: The paper emphasizes that counterfactuals generated through VAEs have "theoretically sound" foundations in latent space representations, suggesting that explanations grounded in formal mathematical principles may be more reliable than ad-hoc explanation methods.