Explainable AI: Techniques for Transparent Decision-Making in Neural Networks

Abstract

Artificial Intelligence (AI), particularly deep neural networks, has achieved strong predictive performance across healthcare, finance, autonomous systems, cybersecurity, and natural language processing, yet its increasing complexity has intensified concerns about transparency and interpretability. This review examines Explainable Artificial Intelligence (XAI) techniques for improving transparent decision-making in neural-network-based systems. It synthesizes major categories of explainability, including intrinsic and post-hoc approaches, and critically reviews LIME, SHAP, Grad-CAM, Integrated Gradients, saliency methods, Layer-wise Relevance Propagation, attention-based explanations, and counterfactual explanations. The study compares these methods in terms of model dependency, explanation scope, computational requirements, architectural compatibility, fidelity, robustness, and interpretability. It also evaluates current approaches using criteria such as faithfulness, sensitivity, stability, human comprehensibility, scalability, and fairness. Applications across healthcare, finance, autonomous systems, cybersecurity, and natural language processing are discussed to demonstrate the practical relevance of explainable neural networks. Key limitations include explanation fragility, adversarial manipulation, computational cost, bias, privacy concerns, and the absence of standardized evaluation frameworks. Finally, the review identifies emerging directions involving XAI 2.0, concept-based explanations, counterfactual recourse, multimodal and generative AI, and human-centered explainability, emphasizing the need for reliable, actionable, and trustworthy neural-network explanations. These directions are essential for responsible deployment and effective human oversight in high-stakes decision environments.