Contrastive Counterfactual Visual Explanations With Overdetermination

Explanations should meet three criteria: The paper argues that good AI explanations need to be contrastive (comparing what changed), counterfactual (showing what would happen if things were different), and measurable—a framework that could help standardize how we evaluate whether AI systems can truly justify their decisions.

Medical imaging explanations improved significantly: CLEAR Image outperformed established explanation methods like Grad-CAM and LIME by an average of 27% on a medical imaging task, suggesting this approach may be particularly valuable for high-stakes healthcare applications where doctors need to understand why an AI flagged a concern.

The method identifies redundant decision-making: CLEAR Image excels at detecting "causal overdetermination"—situations where multiple different parts of an image could independently lead to the same diagnosis. This is important because it reveals when an AI system might be relying on multiple suspicious indicators rather than a single clear finding, which has implications for how much we should trust the classification.

Automatic generation of comparison images: Rather than relying on manual selection or simple masking, CLEAR Image uses adversarial learning to automatically generate appropriate counterfactual images for comparison, potentially making explanations more objective and reproducible across different use cases.