Theory and Methodology for Trustworthy Machine Learning

In recent years, AI and machine learning technologies have advanced at a remarkable pace. They are becoming increasingly embedded in our daily lives, extending to critical areas of societal decision-making, such as medical diagnosis. At the same time, this situation compels us to pay closer attention to a fundamental question:

“Is the output provided by AI truly accurate?”

“To what extent can we trust it?”

These questions have become increasingly pressing in a world where AI delivers outputs with remarkable sophistication and precision—often so convincing that we are tempted to accept it without question. To build a society where AI and humanity can coexist safely, ensuring the trustworthiness of AI systems is no longer optional; it is an essential requirement. But then, what conditions must be satisfied to guarantee such trustworthiness? I explore this question by focusing on Bayesian inference, learning theory, and robust machine learning, from the perspectives of generalization, robustness, and uncertainty quantification. Moreover, I aim to apply the insights gained through this inquiry to various scientific disciplines, ultimately establishing processes for reliable knowledge discovery and trusted decision-making.