Enhancing Explainable AI: A Hybrid Approach Combining GradCAM and LRP for CNN Interpretability

Combining multiple explanation methods can improve AI transparency: The paper demonstrates that merging two existing explanation techniques (GradCAM and LRP) produces better results than using either method alone, suggesting that hybrid approaches may be a practical way to make AI decisions more understandable.

Simpler explanations don't mean less accurate ones: The hybrid method performed better on "Complexity," meaning it provided clearer, less cluttered visual explanations while maintaining or improving performance on other quality measures—demonstrating that interpretability and accuracy aren't necessarily trade-offs.

Visual explanations for AI need rigorous testing: The researchers evaluated their method across five different metrics (Faithfulness, Robustness, Complexity, Localisation, and Randomisation), highlighting that responsible AI requires multiple ways to verify that explanations actually reflect how the model makes decisions.

This work addresses image recognition systems specifically: The focus on CNNs (Convolutional Neural Networks) means these findings apply to visual AI systems like medical imaging, autonomous vehicles, and facial recognition—areas where understanding why the AI decided something matters greatly.

Practical improvements in explainability are achievable: By combining noise removal, element-wise multiplication, and Gaussian blur techniques, the researchers show that meaningful gains in explanation quality can come from relatively straightforward computational adjustments.