Abstract
Interpreting the decision-making process of Convolutional Neural Networks (CNNs) is critical for deploying models in high-stakes domains. Gradient-weighted Class Activation Mapping (Grad-CAM) is a widely used method for visual explanations, yet it typically focuses on the final convolutional layer or naïvely averages across layers, strategies that can obscure important semantic cues or amplify irrelevant noise. We propose Winsor-CAM, a novel, human-tunable extension of Grad-CAM that generates robust and coherent saliency maps by aggregating information across all convolutional layers. To mitigate the influence of noisy or extreme attribution values, Winsor-CAM applies Winsorization, a percentile-based outlier attenuation technique. A user-controllable threshold allows for semantic-level tuning, enabling flexible exploration of model behavior across representational hierarchies. Evaluations on standard architectures (ResNet50, DenseNet121, VGG16, InceptionV3) using the PASCAL VOC 2012 dataset demonstrate that Winsor-CAM produces more interpretable heatmaps and achieves superior performance in localization metrics, including intersection-over-union and center-of-mass alignment, when compared to Grad-CAM and uniform layer-averaging baselines. Winsor-CAM advances the goal of trustworthy AI by offering interpretable, multi-layer insights with human-in-the-loop control.
Introduction
While Convolutional Neural Networks (CNNs) dominate computer vision, they remain "black boxes." To trust them in high-stakes fields like healthcare or autonomous driving, we need to see why they make a specific prediction.
The current gold standard, Grad-CAM, typically generates heatmaps (saliency maps) based only on the final convolutional layer. This creates a "semantic bias": the model shows the general area of an object but often ignores the fine-grained details processed in earlier layers. Conversely, simply averaging all layers together usually results in a noisy output because a few "dominant" layers with extreme activation values drown out the subtle, informative ones. Winsor-CAM aims to bridge this gap by providing a clean, multi-scale explanation that humans can interactively tune.
Methodology
Winsor-CAM operates in a single pass through six main steps:
- Layer-wise Grad-CAM: It calculates a Grad-CAM map for every convolutional layer in the network.
- Spatial Alignment: All maps are upsampled to a common resolution.
- Importance Extraction: A scalar "importance score" is calculated for each layer.
- Winsorization (The Core Innovation): To prevent a few layers from dominating the final output, the authors apply one-sided upper clipping. Using a user-defined percentile parameter $p$, the algorithm clips the highest importance scores to a threshold. This "mutes" the loudest layers, allowing the collective "voice" of the entire network to be heard.
- Normalization: Scores are scaled to a standard range while preserving zero-values.
- Linear Fusion: The final heatmap is a weighted sum of all layer maps based on these Winsorized scores.
The parameter $p$ acts as a "semantic knob": lower values emphasize shallow features (edges, textures), while higher values prioritize deep, abstract representations.
Experimental Setup
The authors evaluated Winsor-CAM across six different CNN architectures (including ResNet, DenseNet, and VGG) using two primary datasets:
- PASCAL VOC 2012: A standard benchmark for natural image classification and object localization.
- PolypGen: A medical imaging dataset for polyp segmentation, used to test the method's generalizability to specialized, high-stakes domains.
Key Results and Discussion
Winsor-CAM consistently outperformed seven major baselines, including Grad-CAM, Grad-CAM++, and FullGrad.
- Localization: On DenseNet121 (PASCAL VOC), Winsor-CAM achieved 46.8% IoU (Intersection over Union), compared to 39.0% for standard Grad-CAM.
- Spatial Alignment: It reduced the "Center-of-Mass" distance (the gap between the heatmap's center and the actual object) from 0.074 to 0.059.
- Fidelity: In "insertion/deletion" tests, Winsor-CAM showed superior faithfulness to the model's actual logic.
Qualitatively, Winsor-CAM provides cleaner heatmaps that are less sensitive to interpolation methods. The "outlier problem" in multi-layer aggregation is effectively mitigated by Winsorization, preventing a few high-gradient layers from masking useful information elsewhere in the network.
Conclusion and Future Work
Winsor-CAM introduces a robust, human-in-the-loop approach to XAI. While the $p$ parameter offers flexibility, future work may explore adaptive Winsorization—where the model automatically selects the best $p$ based on image characteristics. Additionally, the medical imaging evaluation revealed that standard "blurring" baselines used in XAI metrics can sometimes look like medical pathologies, suggesting a need for more domain-specific evaluation protocols in medical XAI. Finally, the authors plan to conduct user studies to see how experts actually interact with the "tunable" aspect of the tool in real-world diagnostic workflows.
Summary updated