Fast Hierarchical Games for Image Explanations.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Complex neural networks achieve high performance but lack interpretability.
• Lack of transparency hinders AI deployment in critical domains like healthcare.
• Current explanation methods struggle with scalability and accuracy.

Purpose of the Study:

• To present Hierarchical Shap (h-Shap), a novel model-agnostic explanation method for image classification.
• To address the limitations of existing interpretability techniques, particularly Shapley-based methods.
• To provide a scalable and computationally efficient approach for explaining AI predictions.

Main Methods:

• Developed a hierarchical extension of Shapley coefficients called Hierarchical Shap (h-Shap).
• Designed h-Shap to be scalable and computable without approximations.
• Leveraged distributional assumptions, common in multiple instance learning, for exact Shapley coefficient retrieval.

Main Results:

• h-Shap demonstrated superior performance compared to state-of-the-art methods in accuracy and runtime.
• The method achieved exponential improvement in computational complexity under specific assumptions.
• Evaluated on synthetic, medical imaging, and general computer vision datasets.

Conclusions:

• Hierarchical Shap (h-Shap) offers a significant advancement in AI model interpretability.
• The method provides an accurate, scalable, and efficient solution for explaining image classification models.
• Publicly available code and experiments facilitate further research and adoption.