Causally-Informed Instance-Wise Feature Selection for Explaining Visual Classifiers.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Black-box image classifiers lack transparency, hindering trust and debugging.
• Existing interpretability methods often fail to capture true causal relationships.

Purpose of the Study:

• To propose a novel interpretability framework for black-box image classifiers.
• To identify input regions with the greatest causal influence on model predictions.

Main Methods:

• Integrating instance-wise feature selection with causal reasoning.
• Formalizing causal influence using a structural causal model and conditional mutual information.
• Employing continuous subset sampling and Rényi's α-order entropy for optimization.

Main Results:

• The proposed method generates compact, semantically meaningful, and causally grounded explanations.
• Experiments show superior performance over existing baselines in predictive fidelity across vision datasets.

Conclusions:

• This framework offers a robust approach to understanding black-box image classifier decisions.
• Causal reasoning provides a more reliable basis for interpretability than traditional feature importance measures.