Patch-Based Explainability for CNN Classification Models.

This work presents a rule-based approach for explaining the predictions of neural network classifiers, with a particular focus on Convolutional Neural Networks (CNNs). The FidexGlo algorithm was used to explain CNN decisions not at the pixel level but through patch-based conditions, where each antecedent expresses the average intensity of an image patch. Experiments were conducted on three standard benchmarks of grayscale images: MNIST, Fashion-MNIST, and EMNIST (letters). A ResNet-50 architecture was fine-tuned for each dataset. FidexGlo was then applied to the average values of image patches to extract rules closely aligned with the CNN’s behaviour. Qualitative visualisations demonstrate that the extracted rules identify meaningful discriminative regions—such as holes in digits, or shape boundaries. Overall, the study shows that our explainability method enables efficient, interpretable rule extraction from CNNs using patch-based explanations, offering a human-understandable view of deep model decisions. FidexGlo is available at https://github.com/Jean-Marc-B/dimlpfidex_Hepia .