Exploring the Divergence Between Traditional Bottom-up and Explainable Artificial Intelligence Saliency Maps for Interpretability

This paper explores the convergence of traditional and explainable AI (xAI) saliency maps to improve interpretability in deep learning models. It compares a classic saliency map approach (VOCUS2) with a current xAI method (Grad-CAM++), analyzing their advantages, disadvantages, and implications in the context of traffic sign detection for autonomous vehicles. The study examines the implementation of both traditional and xAI saliency maps, their shortcomings, and relevant social concerns. It also evaluates how the models perform when faced with different types of noise and variations.