Explaining Convolutional Neural Networks: A Literature Overview and Implementation of Explainability Methods

Convolutional Neural Networks (CNNs) have revolutionized various fields, achieving impressive perfor- mance in tasks such as image classification, object detection, and medical image analysis. Despite their success, the ”black-box” nature of CNNs poses ethical challenges concerning transparency, trust, and accountability in AI systems. Our project aims to address these ethical concerns by providing a compre- hensive literature overview of explainability methods for CNNs and implementing some of these methods in practice using pretrained models such as ResNet-50. Additionally, we will discuss the challenges of feature visualization techniques and attempt to implement them, even if without regularization, to provide further insights.