Winsor-CAM: Human-Tunable Visual Explanations from Deep Networks via Layer-Wise Winsorization

Winsor-CAM is a robust, single-pass visual explanation method that aggregates multi-layer Grad-CAM maps and applies user-controllable percentile-based Winsorization to produce more accurate and tunable saliency maps than existing baselines across diverse CNN architectures and medical imaging tasks.

The Gist

🎨 The Problem: The "One-Layer" Blind Spot

Imagine you are trying to explain to a friend why a computer program recognized a picture of a Golden Eagle.

Most standard tools (like the popular Grad-CAM) act like a single-lens camera. They only look at the very last step of the computer's thinking process.

• The Good: They see the big picture: "Ah, that's a bird!"
• The Bad: They miss the details. They might miss the specific texture of the feathers, the shape of the beak, or the way the light hits the wing because those details were processed in the earlier steps of the computer's brain.
• The Result: The explanation is often blurry, noisy, or highlights the background (like the sky) instead of the bird.

Furthermore, if you tried to just "average" the thoughts from every step of the computer's brain, you'd get a mess. The loudest, most aggressive thoughts (usually from the deep, complex layers) would drown out the quiet, important whispers from the early layers (like edges and textures).

💡 The Solution: Winsor-CAM (The "Smart Committee")

The authors of this paper created Winsor-CAM. Think of this not as a camera, but as a smart committee meeting where every layer of the neural network gets to speak, but with a special rule to keep the meeting productive.

Here is how it works, step-by-step:

1. Gathering the Voices (Multi-Layer Aggregation)

Instead of only listening to the "CEO" (the final layer), Winsor-CAM invites everyone from the "Interns" (early layers that see edges and colors) to the "Managers" (deep layers that see shapes and objects) to the meeting.

• Analogy: It's like asking a whole team of detectives to solve a crime. The rookie sees the muddy footprints; the veteran sees the motive. You need both to get the full story.

2. The "Winsorization" Rule (Taming the Loudmouths)

In any meeting, sometimes one person shouts so loud they drown everyone else out. In neural networks, the deep layers often have huge numbers that overwhelm the early layers.

• The Fix: The paper uses a statistical trick called Winsorization. Imagine a moderator who says: "If anyone's voice is louder than the top 10% of the room, we will cap their volume so they don't dominate the conversation."
• This doesn't silence them; it just turns the volume down to a reasonable level so the quieter, subtle clues from the early layers can still be heard.

3. The "Human Tuner" (The Volume Knob)

This is the coolest part. The user gets a dial (called the percentile parameter, $p$).

• Turn the dial down (Low $p$): You tell the committee, "Ignore the big picture managers; I want to hear the interns." The result? The explanation highlights fine details like fur texture, leaf veins, or the edge of a tumor.
• Turn the dial up (High $p$): You tell the committee, "I want the big picture." The result? The explanation highlights broad shapes like the whole animal or the general organ.
• Why this matters: A radiologist might want to see the texture of a polyp (low $p$), while a model developer might want to see the shape of the object (high $p$). You can adjust the explanation to fit your specific need.

🏆 The Results: Why It Wins

The researchers tested this on six different types of AI brains (like ResNet and DenseNet) using two very different datasets:

1. PASCAL VOC: Pictures of everyday objects (dogs, eagles, cars).
2. PolypGen: Medical images of polyps (tiny growths in the colon).

The Findings:

• Better Focus: Winsor-CAM found the objects more accurately than the old methods. If the old method said "The bird is here," Winsor-CAM said "The bird is right here," with a tighter, cleaner outline.
• Robustness: Even if you picked a "bad" setting for the dial, Winsor-CAM still performed better than the standard methods. It's like a car that drives well even if you forget to adjust the seat; the old cars would crash.
• Medical Magic: In the medical tests, it successfully highlighted the tiny, tricky polyps better than the competition. This is huge because in medicine, missing a tiny detail can be dangerous.

🚀 The Big Picture

Winsor-CAM is like giving a human expert a remote control for the AI's "brain waves."

Instead of getting a static, blurry map that the AI forces on you, you can now:

1. Listen to everyone (all layers of the network).
2. Stop the shouting (suppress the outliers).
3. Tune the focus (zoom in on textures or zoom out on shapes).

It makes AI less of a "black box" and more of a transparent tool that experts can actually trust and use to make better decisions, whether they are identifying a bald eagle or spotting a medical issue.

Technical Summary

1. Problem Statement

Convolutional Neural Networks (CNNs) are critical in safety-sensitive domains like healthcare and autonomous systems, yet their decision-making processes remain opaque. While visual explanation methods like Grad-CAM are widely used, they suffer from significant limitations:

• Single-Layer Limitation: Standard Grad-CAM relies solely on the final convolutional layer, potentially missing crucial low-level cues (textures, edges) learned in earlier layers.
• Instability and Noise: Naïve aggregation of Grad-CAM maps across all layers often introduces noise from less relevant feature maps, diluting semantically meaningful patterns.
• Lack of Control: Existing multi-layer methods (e.g., FullGrad) typically use fixed weighting schemes, offering no mechanism for users to dynamically tune the semantic resolution of the explanation (e.g., shifting focus from fine-grained textures to high-level object shapes).

2. Methodology: Winsor-CAM

The authors propose Winsor-CAM, a single-pass, gradient-based method that aggregates saliency information from all convolutional layers while applying statistical outlier suppression. The pipeline consists of six key steps:

1. Layer-Wise Grad-CAM Computation: Compute standard Grad-CAM maps ($L^c_{i}$) for every convolutional layer $i$ in the network.
2. Spatial Alignment: Upsample all layer-specific maps to a common resolution ($H, W$) using interpolation (e.g., bilinear) to ensure spatial alignment.
3. Importance Score Extraction: Calculate a scalar importance score ($\Gamma^c_i$) for each layer by aggregating the filter-wise weights ($\alpha^c_{i,k}$) using either mean or max pooling, followed by a ReLU activation.
4. Winsorization (Outlier Suppression): This is the core innovation. Instead of using raw scores, the method applies one-sided Winsorization at a user-defined percentile $p$.
   • It computes the $p$-th percentile threshold ($T$) of the non-zero importance scores.
   • Any score $\Gamma^c_i$ exceeding $T$ is clipped to $T$.
   • This suppresses "outlier" layers that might disproportionately dominate the final map due to large activations, while preserving the relative ordering of the remaining layers.
5. Normalization: The clipped scores are min-max normalized to a specific range (e.g., $[0.1, 1.0]$) to ensure bounded contributions, with zero values preserved for layers with no positive influence.
6. Final Fusion: The final heatmap is generated as a weighted linear combination of the interpolated layer maps, using the normalized Winsorized scores as weights.

Key Feature: The percentile parameter $p$ acts as a human-tunable control knob.

• Low $p$: Aggressively clips high scores, emphasizing early-layer features (edges, textures).
• High $p$: Retains more high-layer contributions, emphasizing abstract, high-level semantic patterns.

3. Key Contributions

1. First Multi-Layer Aggregation with Winsorization: Introduces the first method to aggregate Grad-CAM across the entire convolutional stack while using statistical clipping to mitigate outlier dominance.
2. Human-Tunable Semantic Control: Provides a single, interpretable parameter ($p$) that allows users to dynamically adjust the semantic abstraction level of the explanation without retraining the model or performing multiple passes.
3. Comprehensive Evaluation: Extensive benchmarking across six diverse CNN architectures (ResNet50, DenseNet121, VGG16, InceptionV3, EfficientNet-B0, ConvNeXt-Tiny) on two datasets: PASCAL VOC 2012 (natural images) and PolypGen (medical imaging).
4. Robustness: Demonstrates that even the worst-performing fixed $p$-value configuration outperforms state-of-the-art baselines like FullGrad, proving the method's inherent robustness.

4. Experimental Results

The method was evaluated using Intersection over Union (IoU), Center-of-Mass (CoM) distance, Insertion AUC, and Deletion AUC.

Performance on PASCAL VOC 2012 (DenseNet121):

• Localization (IoU): Winsor-CAM achieved 46.8% IoU, significantly outperforming standard Grad-CAM (39.0%) and FullGrad (43.3%).
• Spatial Alignment (CoM Distance): Reduced error to 0.059 (vs. 0.074 for Grad-CAM and 0.072 for FullGrad).
• Fidelity: Improved Insertion AUC (0.656 vs. 0.623) and Deletion AUC (0.197 vs. 0.242).
• Ablation Study: Confirmed that including earlier layers improves localization, provided outlier contributions are suppressed via Winsorization.

Performance on PolypGen (Medical Imaging):

• Winsor-CAM maintained superior localization performance (IoU and CoM distance) across all architectures compared to baselines, despite the challenges of medical image baselines (e.g., polyp-like artifacts in blurred baselines).
• The method proved effective in domain-specific contexts where precise localization is critical for diagnostic support.

Comparison with Baselines:
Winsor-CAM consistently outperformed seven baselines: Grad-CAM, Grad-CAM++, LayerCAM, ScoreCAM, AblationCAM, ShapleyCAM, and FullGrad. Notably, it surpassed FullGrad (which also aggregates all layers) by effectively managing layer weighting through percentile clipping rather than uniform summation.

5. Significance and Impact

• Expert-in-the-Loop Analysis: Winsor-CAM bridges the gap between automated attribution and human interpretability. It allows domain experts (e.g., radiologists) to interactively tune the explanation to focus on specific feature hierarchies relevant to their task.
• Efficiency: Unlike multi-pass methods (e.g., Integrated Gradients, ShapleyCAM) that require repeated forward/backward passes, Winsor-CAM maintains the single-pass efficiency of standard Grad-CAM.
• Robustness: The statistical approach to outlier suppression makes the method robust to architectural variations and specific dataset characteristics, offering a reliable tool for safety-critical applications where trust and transparency are paramount.

In conclusion, Winsor-CAM represents a significant advancement in Explainable AI (XAI) by solving the multi-layer aggregation problem through a statistically grounded, user-controllable mechanism, delivering superior localization and interpretability across diverse domains.