Robust Perturbation for Visual Explanation: Cross-Checking Mask Optimization to Avoid Class Distortion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning Interpretability

Background:

• Deep neural networks (DNNs) achieve high performance but lack transparency.
• Visualizing attribution maps is crucial for understanding DNN decisions in computer vision.
• Perturbation-based methods explain DNNs by optimizing masks to alter predictions, but can lack robustness.

Purpose of the Study:

• To address the issue of class distortion in perturbation-based DNN visualization.
• To propose a novel framework, Robust Perturbation, for reliable and accurate visual explanations.
• To enhance the robustness and fidelity of perturbation mechanisms in DNN interpretation.

Main Methods:

• Defined class distortion as unexpected prediction changes during perturbation.
• Developed Robust Perturbation, a framework employing a cross-checking mask optimization strategy.
• Evaluated the framework on three public datasets and proposed a new metric for class distortion.

Main Results:

• Robust Perturbation demonstrated robustness against class distortion during mask optimization.
• The framework effectively perturbs target predictions while preserving non-target predictions.
• Quantitative and qualitative experiments confirmed enhanced quality and fidelity of visual explanations.

Conclusions:

• Tackling class distortion significantly improves the reliability of DNN visual explanations.
• Robust Perturbation offers a more trustworthy approach to understanding DNN decision-making.
• The proposed framework and metric advance the field of explainable artificial intelligence.