Exploring Explainability and Transparency in Deep Neural Networks: A Comparative Approach
J. Bhagya, Jeena Thomas, Ebin Deni Raj
Abstract
Explainability in Artificial Intelligence (AI) helps define model accuracy, clarity, transparency, and results in decision-making backed by AI. A business needs to establish reliability and trust when implementing AI models. Explainability provides a responsible approach to AI development. This paper’s primary objective is to examine the visual explanation property of various interpretability methods. An attempt is made to understand how a convolutional neural network’s output can be explained. It was common practice to refer to the deep learning models as black boxes; however, our primary concern is determining precisely what takes place within them. Explainability will give us an understanding of which all elements lead to a prediction. A few techniques for providing visual explanations from Convolutional Neural Network (CNN)-based models are the focus of our investigation. The transparency that can be achieved through visualization receives the most attention. Along with this, attention is given to class discrimination and whether the model takes into account the smallest of details. There have been several proposals for creating deep learning models that are intuitive. In this paper, an attempt is made to determine the effectiveness and drawbacks of Grad-CAM, Ablation-CAM, Score-CAM, and Eigen-CAM.