Interpreting the Incomprehensible: Benchmarking Visual Explanation Methods for Deep Convolutional Networks

Wei Hong Lim^*1, Marwa M. Eid²

¹ Faculty of Engineering, Technology and Built Environment, UCSI University,

 Kuala Lumpur 56000, Malaysia

² Faculty of Artificial Intelligence, Delta University for Science and Technology,

 Mansoura 11152, Egypt

Emails limwh@ucsiuniverisity.edu.my; mmm@ieee.org

Abstract

Deep Convolutional Networks (CNNs) have revolutionized various fields, including computer vision, but their decision-making process remains largely opaque. To address this interpretability challenge, numerous visual explanation methods have been proposed. However, a comprehensive evaluation and benchmarking of these methods are essential to understand their strengths, limitations, and comparative performance. In this paper, we present a systematic study that benchmarks and compares various visual explanation techniques for deep CNNs. We propose a standardized evaluation framework consisting of benchmark explain ability methods. Through extensive experiments, we analyze the effectiveness, and interpretability of popular visual explanation methods, including gradient-based methods, activation maximization, and attention mechanisms. Our results reveal nuanced differences between the methods, highlighting their trade-offs and potential applications. We conduce a comprehensive evaluation of visual explanation methods on different deep CNNs, the results demonstrate the ability to achieve informed selection and adoption of appropriate techniques for interpretability in real-world applications.

Keywords: Convolutional Neural Networks (CNNs); Benchmarking, Interpretability; Class Activation Maps (CAM); Deep learning, Image classification; Explainable AI.