Statistical Reliability of Data-Driven Science and Technology

With the rapid development of AI and machine learning, the use of data-driven approaches has been expanding across various fields of science and technology. In data-driven approaches, unlike traditional scientific research and technological development, hypotheses are generated based on data, requiring the consideration of data dependency when evaluating hypotheses. As a result, conventional statistical tests, which have served as the foundation for reliability assessments in scientific research and technological development, are inadequate for properly evaluating the reliability of data-driven hypotheses. In this paper, we introduce the framework known as selective inference, which has gained attention as a statistical reliability evaluation method for data-driven science and technology. We provide an overview of recent research trends in selective inference and present our recent studies on statistical tests for deep learning models based on selective inference. (c) 2025 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.