SketchXplain: Intuitive Visual Explanations of Image Classifiers with Sketches

Current AI explanations have a significant interpretability gap: Traditional saliency maps (heat maps showing which image regions influenced a prediction) often fail to be intuitive or semantically clear to users, meaning people struggle to understand why an AI made a decision even when shown which parts of an image mattered.

Sketch-based explanations enable faster and more accurate interpretation: User studies demonstrated that SketchXplain's sketch-based visual explanations supported "quicker interpretation with more aligned visualizations than saliency maps or simple drawings," suggesting that how we present AI reasoning matters as much as what information we show.

Simplicity and selectivity improve AI transparency: By combining saliency detection with concept-based models and artistic abstraction, SketchXplain demonstrates that removing unnecessary visual noise and focusing on key features makes AI decisions more understandable—a principle applicable beyond just image classification.

Sketch explanations work across different domains: The approach proved effective for both face expression recognition and medical diagnosis (skin lesion detection), with the medical application showing that sketches "more coherently visualized disease symptoms" and better supported non-expert understanding of diagnostic decisions.

Responsible AI requires human-centered design of explanations: This work illustrates that explainable AI isn't just about technical accuracy—it requires designing explanations that align with how humans naturally understand information, drawing inspiration from how people communicate visually through sketches.