EXPLAINABLE AI TECHNIQUES FOR IMPROVING TRUST IN DEEP LEARNING MODELS

Abstract

Deep learning models have a major obstacle of the lack of insight into the reasons behind the choices of the model that could help in their use in the high-stakes settings, including healthcare and finance, where predictive accuracy isn’t the most important (but rather just one of the two). In this paper, we provide an in-depth analysis of methods for Explainable AI (XAI) to promote trust and transparency in deep learning systems. Our comparative study of four leading XAI approaches, such as Grad-CAM, LIME, SHAP, and Saliency Maps, is carried out in various aspects, such as performance, interpretability, trustworthiness, and computational efficiency. Our approach is to use standardized measures of evaluation on three varying datasets (CIFAR-10, Lending Club, and Chest X-Ray) that contain 480 participants who are experts and non-experts. The findings indicate that Grad-CAM is the best in terms of balance, 85.1% accuracy, 0.012s explanation time, whereas SHAP performs better in terms of faithfulness (0.891) in explaining model decisions. Markedly, we also find that the domain experts give a higher rating in Grad-CAM (4.35/5 trust score) when it comes to visual tasks, whereas non-experts give preference to LIME (4.26/5) due to its explanations of feature importance based on intuitions. Regarding statistical analysis, our data proves the presence of significant performance differences between the methods (p < 0.001) with significant effect sizes (Cohen's d > 0.85). As demonstrated in the real world, XAI integration is found to decrease decision time by 42.3% in healthcare and regulatory compliance by 94.2% in finance. The study delivers a proven framework of selecting the right XAI methods depending on the requirements of a particular application, and this will help introduce more transparent, trustworthy, and deployable AI systems to critical fields. Its implications include the fact that XAI needs to be selected contextually, as this is the only way to establish human trust in the use of AI in decision-making.