Organization・Member

Over the years, humanity has discovered the fundamental laws and basic equations of physics through observations and experiments on natural phenomena and deep insights of scientists.

Computational physics takes a deductive approach to elucidate and predict complex physical phenomena based on these fundamental equations using computer simulations.

Machine learning/AI, on the other hand, is an inductive approach that uses data from observations, experiments, or computer simulations to predict unknown data without relying on physical laws.

The Intelligence Learning team is conducting research that bridges data science and conventional physics. This includes extracting parameters that enhance the understanding of large datasets by leveraging the insights gained from statistical physics and condensed matter physics.

In addition, the team addresses previously challenging problems with many degrees of freedom by combining experimental and observational data with computer simulations.

We aim to bridge the gap between data science and conventional physics.

In physics, machine learning is mainly used to estimate physical quantities from large amounts of experimental data and to extract features of the data. The data used in physics are of high quality, with errors and reliability controlled. In some cases, even the results of theoretical calculations are used as data to identify phase diagrams or quantum states.

For these purposes, it is important to design learning algorithms that not only have excellent predictive performance, but also provide an assessment of reliability.

Working with high-quality data from physics is also useful for studying theoretically how the learning process proceeds in an idealized environment. Machine learning with neural

networks often shows interesting parallels to the function of neurons in the brain.

Theoretical studies of machine learning, in turn, are expected to lead to a mathematical representation of how our brain works and to the development of mathematical physics

that can elucidate its complex processes, i.e., the emergence of intelligence.

Multilayer neural networks used in deep learning are expected to be ultimately understood as nonlinear many-body systems. At the Institute for Physics of Intelligence, we aim to explore the manifestation of intelligence from the theoretical perspectives of nonequilibrium statistical mechanics and condensed matter physics.

By combining cutting-edge physics theory with innovative experiments in nonlinear and nonequilibrium physics,we will develop new methodologies to understand the learning process in deep learning and enhance its efficiency.

In addition to addressing fundamental principles in deep learning, we also aim to apply it to the understanding of various phenomena in condensed matter physics, biophysics, and astrophysics.

Our goal is to accelerate scientific discovery in a wide range of physics fields through machine learning and deep learning, with the aim of eventually building machines that can make discoveries on their own.。