Improving Deep Learning Transparency: Leveraging the Power of LIME Heatmap

Deep learning techniques have recently demonstrated remarkable precision in executing tasks, particularly in image classification. However, their intricate structures make them mysterious even to knowledgeable users, obscuring the rationale behind their decision-making procedures. Therefore, interpreter methodologies have emerged to introduce clarity into these techniques. Among these approaches is the Local Interpretable Model-Agnostic Explanations (LIME), which stands out as a means to enhance comprehensibility. We believe that interpretable deep learning methods have unrealised potential in a variety of application domains, an aspect that has been largely neglected in the existing literature. This research aims to demonstrate the utility of features like the LIME heatmap in advancing classification accuracy within a designated decision-support framework. Real-world contexts take centre stage as we illustrate how the heatmap determines the image segments playing the greatest influence on class scoring. This critical insight empowers users to formulate sensitivity analyses and discover how manipulation of the identified feature could potentially mislead the deep learning classifier. As a second significant contribution, we examine the LIME heatmap data of GoogLeNet and SqueezeNet, two prevalent network models, in an effort to improve the comprehension of these models. Furthermore, we compare LIME with another recognised interpretive method known as Gradient-weighted Class Activation Mapping (Grad-CAM), evaluating their performance comprehensively. Experiments and evaluations conducted on real-world datasets containing images of fish readily demonstrate the superiority of the method, thereby validating our hypothesis.