DeDUCE: Generating Counterfactual Explanations Efficiently

Counterfactual explanations help identify what would need to change for AI to make correct decisions: When image classifiers make mistakes, showing users what modifications to an image would lead to the correct classification provides actionable insight into how the AI system works, supporting more transparent AI deployment.

The new DeDUCE algorithm is computationally efficient and scalable: Unlike previous methods, this approach can work with large image classifiers at low computational cost, making counterfactual explanations practical for real-world applications rather than just research settings.

DeDUCE generates more realistic and minimal changes compared to existing methods: The algorithm consistently finds counterfactuals that are "much closer to the original inputs" while maintaining comparable realism to baseline approaches, meaning users see smaller, more believable modifications needed for correct classification.

The method works specifically with spectral normalization, a common AI safety technique: By designing the algorithm to work with spectral normalization—a regularization method used to improve neural network robustness—the approach integrates with existing responsible AI practices rather than requiring separate systems.

Open-source availability enables broader adoption and verification: Publishing code publicly allows other researchers and practitioners to test, validate, and build upon the method, supporting reproducibility and trust in AI explanation tools.