설명가능한 인공지능(XAI) 방법론의 산업별 적용가능성에 관한 연구

Abstract

The learning performance of artificial intelligence (AI) technologies such as machine learning and deep learning is approaching or surpassing that of humans, and humans are gaining new insights through hidden patterns and rules discovered by AI, but their delivery and explanatory power is in short supply. As the use of AI technology expands by industry, values such as transparency, fairness, and accountability are continuously required in addition to accuracy. Accordingly, the demand and necessity for eXplainable Artificial Intelligence (XAI) has recently been emphasized. XAI is an analysis model, process, and service for determining the reliability of AI results by explaining what the reason is if there is an error in the output produced by the AI model. This study reviewed SHAP, LIME, and LRP, which are XAI methodologies, and explored the applicability of these methodologies to each industry. In the financial field, algorithms such as SHAP, LIME, and LRP can be applied to support credit rating prediction, loan decision-making, and investment decision-making, and to ensure their explainability and interpretability. This can increase fairness and reliability of AI results between financial consumers and financial institutions. In the defense and military fields, AI staff is being used as an auxiliary tool for key decision makers. If the explainability and interpretability of the results are guaranteed by applying XAI, the role of AI staff will change to active utilization. In the field of industrial accidents, SHAP can be used in a model that predicts industrial accidents to determine the degree of influence of variables that affect the occurrence of industrial accidents. As such, the XAI methodology can contribute to increasing the transparency and interpretability of models in various AI-based prediction models.