SketchXplain uses sketches to explain AI image classifiers more intuitively
New sketch-based XAI method beats saliency maps for faster, clearer AI explanations
Deep Dive
A paper on arXiv introduced SketchXplain, an XAI method that generates sketch-based visual explanations for image classifiers. By combining saliency maps, concept-bottleneck models, and sketch optimization, SketchXplain selects coherent observation artifacts and represents them as simple sketches. User studies on facial expression recognition and skin lesion diagnosis showed SketchXplain supported quicker, more aligned interpretation than saliency maps or simple drawings, and more coherently visualized disease symptoms to better support lay diagnosis.
Key Points
- Combines saliency maps, concept-bottleneck models, and sketch optimization for coherent, simple explanations
- User studies on facial expression recognition showed quicker interpretation and higher alignment than saliency maps
- Evaluated on skin lesion diagnosis: more coherently visualized symptoms, better supporting lay diagnosis
Why It Matters
Sketch-based XAI could make AI decisions transparent for non-experts in critical fields like healthcare.