Technical Summary: SketchXplain

Problem Statement

Current Explainable AI (XAI) methods for image classification, particularly saliency maps, suffer from an "interpretability gap." While saliency maps faithfully highlight relevant pixels or neurons, they often lack semantic context, appearing as unintuitive "nebulous blobs" that fail to convey why a region is relevant. This forces users to expend significant cognitive effort to interpret the model's behavior, hindering quick and accurate decision-making. Existing concept-based explanations improve semantic alignment but often lack visual coherence with the specific image instance, while simple line drawings or edge detection fail to capture the necessary explanatory cues. The authors argue that AI explanations should be intuitive, defined by three desiderata: Simplicity (clear and concise), Coherence (aligned with user knowledge and evidence), and Quick Interpretation (enabling rapid assimilation).

Methodology: SketchXplain

The authors propose SketchXplain, a deep neural network framework that generates sketch-based visual explanations. Unlike methods that simply convert photos to sketches or highlight pixels, SketchXplain integrates saliency, concept bottlenecks, and sketch optimization to produce abstract, cue-aware drawings.

The architecture consists of three main modules:

1. Base Prediction: An image encoder ($M$) predicts the class label ($\hat{y}$) from the input image ($x$).
2. Concept Explanations:
   • Concept Bottleneck: The model predicts intermediate, human-interpretable concepts ($\hat{\gamma}$) (e.g., facial Action Units or skin lesion ABCD criteria) via multi-label binary classification. These concepts are then used to predict the final class label.
   • Cue Localization: Using Grad-CAM adapted for transformers, the system generates saliency maps for each predicted concept. These maps are weighted by the importance of each concept to the final prediction ($\hat{w}_\gamma$) to create a unified, concept-weighted saliency map ($\hat{\epsilon}$).
3. Sketch Rationalization: This module generates the final sketch through two steps:
   • Stroke Initialization: A pretrained image-to-image generator extracts detailed lines (edges, wrinkles, textures) from the photo. These lines are masked by the concept-weighted saliency map to retain only the most explanatory strokes.
   • Stroke Optimization: The system initializes a set of parametric strokes (quadratic Bézier curves) based on the masked lines. These strokes are iteratively optimized via gradient descent to minimize a composite loss function ($L$):
     $$L = -\cos(\hat{z}_s, \hat{z}_x) - \lambda_t \cos(\hat{z}_s, \hat{z}_y) + \lambda_\kappa \kappa(\hat{s}_x)$$
     Where:
     • $\hat{z}_s, \hat{z}_x, \hat{z}_y$ are CLIP embeddings of the sketch, input photo, and predicted label, respectively.
     • The first term ensures Photo Alignment (fidelity to the image).
     • The second term ensures Semantic Alignment (fidelity to the prediction label).
     • The third term ($\kappa$) enforces Smoothness (simplicity) by penalizing high curvature.

Key Contributions

1. Novel Modality: This is the first work to utilize sketches as an intuitive explanatory modality for visual XAI, moving beyond simple attribution maps or verbal concept descriptions.
2. SketchXplain Framework: A technical method that satisfies the desiderata of simplicity, coherence, and quick interpretation by combining concept bottlenecks, saliency localization, and differentiable sketch rendering.
3. Empirical Validation: Comprehensive evaluation across two distinct domains:
   • Facial Expression Recognition: Using Action Units (AUs) as concepts.
   • Skin Lesion Diagnosis: Using ABCD criteria (Asymmetry, Border, Color, Diameter) as concepts.

Results

Modeling and Proxy Metrics

• Simplicity: SketchXplain sketches demonstrated significantly lower visual complexity (measured by local entropy and JPEG XL file size) compared to saliency maps and "Salient Photos," while retaining more information than simple outlines.
• Coherence: SketchXplain achieved the highest alignment with both AI prediction labels and explanatory concepts (AUs/ABCDs) compared to baselines like CLIPasso and Saliency maps. It successfully preserved task-relevant cues while reducing noise.

User Studies (Facial Expressions)

• Qualitative: Participants found SketchXplain more intuitive than saliency maps (which lacked semantics) and simple outlines (which lacked detail). Users reported that SketchXplain effectively highlighted "salient lines" (like Saliency) while maintaining recognizable facial features (like Outlines) and expressive cues (like CLIPasso).
• Quantitative (Time-Constrained): In rapid perception tests (33.3ms – 500ms display durations), SketchXplain enabled significantly higher AI Alignment (correctly identifying the predicted emotion) and Concept Recall (identifying specific AUs) compared to Saliency maps, CLIPasso, and even the original photo at very short durations. Saliency maps performed the worst, likely due to a lack of featural context.

User Studies (Skin Lesions)

• Qualitative: In a study with non-clinical participants, SketchXplain outperformed other visualizations in helping users identify ABCD criteria and diagnose melanoma risk. Participants noted that the "disconnected lines" and "emphasized scribbles" in SketchXplain helped confirm border irregularity and asymmetry more effectively than smooth outlines or blurred saliency masks.
• Trade-offs: While SketchXplain improved interpretability, it introduced slightly higher cognitive load than simple outlines, though users found the trade-off worthwhile for the increased coherence with clinical concepts.

Significance and Claims

The paper claims that SketchXplain addresses the interpretability gap by shifting XAI from purely analytical auditing toward intuitive interpretation. By leveraging the human ability to rapidly process abstract sketches (System 1 thinking), the method supports:

• Quicker Interpretation: Users can assimilate relevant cues and infer predictions faster than with traditional saliency maps.
• Enhanced Coherence: The sketches bridge the gap between spatial geometry and semantic concepts, aligning AI logic with human observation.
• Privacy Preservation: As a byproduct, the abstraction inherent in sketches can protect identifiable information (e.g., identity, gender) while retaining diagnostic utility.

The authors position SketchXplain not as a replacement for all XAI methods, but as a complementary approach that leverages visual abstraction to make AI decisions more accessible to lay users, particularly in domains where rapid, intuitive understanding is critical.