Integrated feature analysis for deep learning interpretation and class activation maps

The paper presents an integrated feature analysis method for deep learning interpretability, particularly focusing on class activation maps (CAMs). It addresses CAM limitations by incorporating feature distribution analysis and feature decomposition to improve consistency, reveal insights into overfitting, dataset characteristics, and feature redundancies. The method enhances CAMs by introducing a common intensity scale, measuring feature importance, and selecting key features. Experiments on eight diverse datasets show significant improvements in consistency, with up to a 64.2% increase, and demonstrate that even a smaller number of features can generate informative saliency maps without compromising performance. The study contributes to a better understanding of model behavior and extends existing CAM algorithms, while also highlighting the need for further research on adapting the approach to other model architectures.