SketchXplain: Intuitive Visual Explanations of Image Classifiers with Sketches
Researchers have developed SketchXplain, a novel method for generating sketch-based visual explanations for image classifiers. This approach aims to bridge the interpretability gap left by traditional saliency maps, which are often unclear. By combining saliency maps with concept-bottleneck models and sketch optimization, SketchXplain selects key visual elements, represents them with concepts, and abstracts them for simplicity. User studies indicate that SketchXplain facilitates quicker and more aligned interpretations compared to existing methods, proving effective in domains like facial expression recognition and skin lesion diagnosis. AI
IMPACT Introduces a novel approach to AI explainability, potentially improving user understanding of image classifier decisions.