Research・Education

The data studied at the Institute for Physics of Intelligence (IPI) are fundamentally different from the large-scale datasets commonly handled in artificial intelligence, in that they are generated by underlying physical principles. In elementary particle physics, observational data arise from fundamental symmetries; in condensed matter physics, from symmetry breaking; and in astrophysics, from structured objects and their collective dynamics formed through cosmic evolution. By contrast, natural language, speech, images, and information on the internet are data generated by the human brain, a highly complex information-processing system.
At IPI, we study these data—whose generative principles are intrinsically different—in a cross-disciplinary and integrative manner. By elucidating the structures and constraints inherent in each type of data from a physical perspective, we aim not only to extend the existing framework of artificial intelligence research but also to expand the frontiers of physics itself. Through such research activities, the Institute is also committed to fostering researchers who can survey and connect fundamental theory and practical applications.

In recent years, artificial intelligence has achieved remarkable performance in prediction, generation, and decision-making through learning with large-scale models and datasets, giving rise to new paradigms such as generative AI and foundation models. At the same time, fundamental questions remain insufficiently understood: by what principles should learning processes and model behaviors be interpreted, and under what conditions can reliable predictions and inferences be guaranteed?

At IPI, we approach learning and inference as non-equilibrium dynamical processes and seek a principled understanding of artificial intelligence using the frameworks of mathematical physics, including stochastic processes, statistical mechanics, and information theory. By theoretically clarifying model expressivity, generalization performance, uncertainty, and fundamental limits of prediction, we aim to construct theories that go beyond empirical performance evaluation and are both testable and predictive. Such physics-based approaches are expected to provide a foundation for understanding the reliability and limitations of artificial intelligence, while simultaneously contributing to deeper insights into natural phenomena and the creation of new knowledge.

Development of New Education and Research
Interdisciplinary of Mathematics and Informatics

Basic Reserch on Physics of Intelligence

• Elucidating the mathematical and physical structures that underpin intelligence
• Application of artificial intelligence and deep learning to the physical world

Education and Human Resource Developmwnt

• Developing human resources who can oversee everything from basic learning to application
• Developing human resources to lead advanced information processing innovation