Transforming gradient-based techniques into interpretable methods
This addresses the challenge of improving interpretability for users of CNNs in image analysis, though it appears incremental as it builds on existing gradient-based techniques.
The paper tackles the problem of noisy visual explanations from gradient-based interpretability methods like Integrated Gradients in CNNs, and introduces GAD (Gradient Artificial Distancing) to reduce image noise and highlight influential regions, with empirical tests on occluded images showing that these regions are pivotal for class differentiation.
The explication of Convolutional Neural Networks (CNN) through xAI techniques often poses challenges in interpretation. The inherent complexity of input features, notably pixels extracted from images, engenders complex correlations. Gradient-based methodologies, exemplified by Integrated Gradients (IG), effectively demonstrate the significance of these features. Nevertheless, the conversion of these explanations into images frequently yields considerable noise. Presently, we introduce GAD (Gradient Artificial Distancing) as a supportive framework for gradient-based techniques. Its primary objective is to accentuate influential regions by establishing distinctions between classes. The essence of GAD is to limit the scope of analysis during visualization and, consequently reduce image noise. Empirical investigations involving occluded images have demonstrated that the identified regions through this methodology indeed play a pivotal role in facilitating class differentiation.