Investigating the Relationship Between Debiasing and Artifact Removal using Saliency Maps

Debiasing works by redirecting model attention: The paper demonstrates that successful debiasing methods systematically shift where neural networks focus their attention, moving focus away from protected attributes (like race or gender). This provides concrete evidence for how fairness interventions actually function at a technical level.

Artifact removal and fairness are connected: Techniques originally designed to remove visual artifacts or irrelevant features in images can be repurposed to improve fairness by eliminating reliance on protected attributes. This suggests that the underlying mechanisms for both goals overlap significantly.

Saliency maps can measure fairness progress: The researchers developed new explainability-based metrics using saliency maps (visual representations of what a model "looks at") to track whether a model's decision-making process has actually shifted during debiasing. This provides a practical tool for monitoring whether fairness interventions are working as intended.

Fairness and artifact removal have a two-way relationship: The paper establishes bidirectional connections between these two concepts, meaning insights from artifact removal research can inform fairness work, and vice versa. This cross-pollination could accelerate progress in both areas.

Explainability tools are essential for fairness: By using explainability techniques to investigate debiasing, the paper highlights that understanding why models make decisions is crucial for building fair AI systems, not just measuring whether they're fair.