FM-G-CAM: Holistic Explainable AI for Computer Vision
A new method called Fused Multi-class Gradient-weighted Class Activation Map (FM-G-CAM) addresses limitations in explainable AI for computer vision. Existing techniques like Grad-CAM focus on a single target class, neglecting much of the prediction process. FM-G-CAM considers multiple top-predicted classes to provide a holistic explanation of a CNN's rationale. The paper includes a detailed mathematical and algorithmic description, along with a comparison to existing methods.
Key facts
- FM-G-CAM stands for Fused Multi-class Gradient-weighted Class Activation Map.
- It improves upon Grad-CAM by considering multiple top-predicted classes.
- The method provides a holistic explanation of CNN predictions.
- The paper includes mathematical and algorithmic descriptions.
- It compares FM-G-CAM with existing explainability methods.
- The research is published on arXiv with ID 2312.05975.
- The announcement type is replace-cross.
- The paper emphasizes the need for explainability in AI for real-world impact.
Entities
Institutions
- arXiv