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

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Interpreting Convolutional Neural Networks (CNNs) is crucial for safety-critical domains like healthcare and autonomous driving.
• Existing methods like Grad-CAM often rely on single layers, potentially missing multi-scale information and leading to unstable explanations.
• There is a need for robust and tunable methods to enhance the reliability of visual explanations for CNNs.

Purpose of the Study:

• To introduce Winsor-CAM, a novel single-pass gradient-based method for interpreting CNNs.
• To enhance the stability and semantic-level tunability of saliency maps generated by CNNs.
• To provide a more comprehensive and robust visual explanation tool for expert-in-the-loop analysis.

Main Methods:

• Winsor-CAM aggregates Grad-CAM maps from all convolutional layers.
• It applies percentile-based Winsorization to reduce the impact of outlier features.
• A user-controllable parameter $p$ allows tuning explanations from low-level textures to high-level object features.

Main Results:

• Winsor-CAM demonstrated superior performance in localization (IoU, CoM distance) and fidelity (insertion/deletion AUC) compared to seven baseline methods across six CNN architectures.
• On DenseNet121, Winsor-CAM achieved 46.8% IoU and 0.059 CoM distance, outperforming Grad-CAM (39.0% IoU, 0.074 CoM distance).
• The method showed effectiveness in both general computer vision tasks (PASCAL VOC 2012) and medical imaging (PolypGen polyp segmentation).

Conclusions:

• Winsor-CAM offers an efficient, robust, and human-tunable approach to CNN interpretation.
• The method's ability to incorporate multi-scale cues and attenuate outliers leads to more stable and reliable saliency maps.
• Winsor-CAM is a valuable tool for enhancing expert-in-the-loop analysis in safety-sensitive applications.