Explainable artificial intelligence (XAI) aims to make discoveries and output generated by machine learning algorithms accessible to human users in a trustworthy and easy-to-understand manner. It describes the impact of an AI model and any potential biases, helping in characterizing models with accuracy, fairness, transparency, and outcomes in decision-making powered by AI. XAI focuses on understanding the data used to train the model, its predictions, the specific elements involved in producing those predictions, and the role of the algorithms used. There are several primary categories of approaches to explain AI systems, such as self-interpretable models and post-hoc justifications. The four pillars of explainable artificial intelligence are explicitness, meaningfulness, explanatory accuracy, and knowledge boundaries. AI has been widely used in the medical field, with applications ranging from clinical decision support to cancer localization in MRI scans to report reading and analysis in radiology and pathology. However, AI has also been applied in other areas of the medical industry, with concerns about its perceived "black box" nature, which can lower trust in its accuracy. This has led to the popularity of explainable artificial intelligence (XAI) in the healthcare industry. This study provides an introductory overview of various methods for explaining and understanding predictions generated by complex machine learning models like deep neural networks. It explores various methodologies, including LIME, anchors, Graph LIME, LRP, DTD, PDA, TCAV, XGNN, SHAP, ASV, break-down, Shapley Flow, textual explanations of visual models, causal models, integrated gradients, meaningful perturbations, and X-NeSyL.
Related articles
Other papers on this site that share keywords with this one, closest match first.
International Journal of Software Computing and Testing
