MI CAM: Mutual Information Weighted Activation Mapping for Causal Visual Explanations of Convolutional Neural Networks

2025-07-11

AI Summary

Convolutional neural networks (CNNs) lack causal interpretability in safety-critical domains such as healthcare and energy. Method: We propose MI-CAM, a mutual information-weighted class activation mapping method that dynamically fuses multi-layer feature maps by quantifying mutual information between feature maps and input images, generating causally plausible saliency maps; causal validity is further verified via counterfactual analysis. Contribution/Results: Unlike existing attribution methods, MI-CAM requires no additional training and operates as a post-hoc, computationally efficient, and theoretically grounded framework. On multiple benchmark datasets, MI-CAM yields qualitatively more human-aligned visualizations and achieves statistically significant improvements in quantitative faithfulness metrics—including Insertion AUC, Deletion AUC, and Faithfulness—outperforming Grad-CAM, Score-CAM, and other state-of-the-art approaches. Notably, it demonstrates superior bias mitigation and decision transparency in fine-grained classification tasks.

Technology Category

• Computer Vision: Interpretability, Explainability, and Transparency
• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods

Abstract

With the intervention of machine vision in our crucial day to day necessities including healthcare and automated power plants, attention has been drawn to the internal mechanisms of convolutional neural networks, and the reason why the network provides specific inferences. This paper proposes a novel post-hoc visual explanation method called MI CAM based on activation mapping. Differing from previous class activation mapping based approaches, MI CAM produces saliency visualizations by weighing each feature map through its mutual information with the input image and the final result is generated by a linear combination of weights and activation maps. It also adheres to producing causal interpretations as validated with the help of counterfactual analysis. We aim to exhibit the visual performance and unbiased justifications for the model inferencing procedure achieved by MI CAM. Our approach works at par with all state-of-the-art methods but particularly outperforms some in terms of qualitative and quantitative measures. The implementation of proposed method can be found on https://anonymous.4open.science/r/MI-CAM-4D27

Problem

• Explains CNN decisions using mutual information-weighted activation maps
• Provides causal visual interpretations via counterfactual analysis validation
• Outperforms existing methods in qualitative and quantitative metrics

Innovation

• Uses mutual information for activation weighting
• Generates causal visual explanations
• Combines weights and activation maps linearly