Artificial Intelligence to Improve Interpretability of Neural Networks: A Comparative Study of Attention Mechanisms, SHAP, and LIME
Authors: Mr. Shriyans Bandebuche Indala College of Engineering [1]
Guide: Mrs. Kalpana Bandebuche [2] Assistant Professor, V.K.K. Menon College, Bhandup (East)
Abstract :
While neural networks are incredibly powerful, their frustrating "black box" problem still holds them back, especially in high-stakes fields where we need to trust the AI's decisions. In this study, we tackle this issue head-on by putting three leading interpretability techniques—attention mechanisms, SHAP, and LIME— to the test to see how they can help us crack open the black box. Our research evaluates these methods across multiple benchmark datasets including CIFAR-10, IMDB, and the UCI Heart Disease dataset. We propose an innovative hybrid framework that effectively balances interpretability requirements with performance objectives. The findings quantitatively demonstrate that attention mechanisms provide exceptional transparency for sequential data processing, while SHAP and LIME offer robust post-hoc explanation capabilities for various model architectures. Our evaluation shows a significant improvement in user trust and understanding, with a measured 40% reduction in false positive analysis time for security experts using SHAP explanations. We also address significant challenges including computational overhead and model transparency issues, while outlining promising future directions for the evolving field of explainable AI (XAI)
Keywords: Interpretability, XAI, Neural Networks, Attention Mechanisms, SHAP, LIME, Model Transparency, Explainable Artificial Intelligence
