Welcome to Takaharu Yaguchi's web page!

 

Professor, Department of Mathematics, Graduate School of Science, Kobe University

JST CREST "Structure-Preserving System Modeling and Simulation Basis Based on Geometric Discrete Mechanics"

1-1 Rokkodai-cho, Nada-ku, Kobe, 657-8501

yaguchi@pearl.kobe-u.ac.jp

Research

I am addressing various real world problems through a mathematical scientific approach. In particular, I am conducting research on simulation and modeling from both theoretical and practical aspects.

Practice: Simulation and Mathematical Modeling

  • Social Network Analysis
    Placeholder image

    With the recent development of sensor devices, it has become possible to acquire a variety of data. I am particularly interested in statistical methods for data obtained as networks in psychology, agriculture and economics.

  • Morphological Computing
    Placeholder image

    In recent years, it has been recognized that a variety of functionalities can be realized only by the movement of the body, as typified by the passive walk of a robot. I am studying these problems from modeling and simulation approaches.

  • Physical Simulation of Musical Instruments
    Placeholder image

    I am conducting research to simulate the sound of a musical instrument by physical simulations of the movement of piano strings, hammers and other parts of the instrument, thereby developing novel physics-based electronic instruments.

Theory: Methods of Mathematical Science

  • Geometric Discrete Mechanics
    Placeholder image

    In order to develop a computer-based physics simulation method, I am working on the physics of the digital world.

  • Model Analysis Using Differential Algebra
    Placeholder image

    I am studying input-output relations using differential algebra for the estimation and analysis of model parameters for simulation.

  • Time Series Models on Statistical Manifolds
    Placeholder image

    Traditionally, the target of time-series analysis has been numerical data. I am developing a time-series analysis method for non-numerical time-evolution data (e.g. evolutional networks) using the information geometry.

Scientific Machine Learning

With the recent development of machine learning techniques such as deep learning, researches to integrate such techniques with scientific computing are in progress. Such integration will enable us to perform scientific computations based on models that are more compatible with data than ever before, and to perform parallel computing using GPUs, which are widely used in the field of deep learning.

Based on the above, we are developing a modeling and simulation platform that enables us

  • to handle phenomena that have been difficult to model,
  • to create models and simulation codes while maintaining laws of physics,
  • and to simplify programming
by combining various knowledge and techniques in informatics and mathematical sciences.

Selected News


Dec. 1, 2021
The following work is accepted for AAAI2022

Y. Chen, T. Matsubara, and T. Yaguchi, “KAM Theory Meets Statistical Learning Theory: Hamiltonian Neural Networks with Non-Zero Training Loss,” AAAI, 2022. (Oral, acceptance rate 4.3%)

Sep. 29, 2021
The following papers are accepted for NeurIPS2021

T. Matsubara, Y. Miyatake, and T. Yaguchi, “Symplectic Adjoint Method for Exact Gradient of Neural ODE with Minimal Memory,” NeurIPS, 2021. (Poster, acceptance rate 26%)

Y. Chen, T. Matsubara, and T. Yaguchi, “Neural Symplectic Form: Learning Hamiltonian Equations on General Coordinate Systems,” NeurIPS, 2021. (Spotlight, acceptance rate 3%)

Oct. 31, 2020
The following work is accepted for the workshop at NeurIPS2020: the Machine Learning and the Physical Sciences 2020.

Shunpei Terakawa, Takashi Matsubara, Takaharu Yaguchi, The Error Analysis of Numerical Integrators for Deep Neural Network Modeling of Differential Equations, the Machine Learning and the Physical Sciences (Workshop at NeurIPS 2020.)

Oct. 9, 2020
The following paper, which proposes a method for modeling physical phenomena using deep learning to construct discrete models with exact energy conservation or dissipation laws, has been accepted as an oral presentation at Neural Information Processing Systems (NeurIPS), a leading conference in the field of machine learning (acceptance rate 1900/9454=20%, oral 105/9454=1.1%)．

Takashi Matsubara, Ai Ishikawa, Takaharu Yaguchi, Deep Energy-based Modeling of Discrete-Time Physics, Advances in Neural Information Processing Systems (NeurIPS), 2020.

Feb. 11, 2020
The following paper has been accepted.

M. Komatsu, S. Terakawa and T. Yaguchi, Energetic-Property-Preserving Numerical Schemes for Coupled Natural Systems, Mathematics.

Oct. 1, 2019
The following research project starts. JST CREST "Structure-Preserving System Modeling and Simulation Basis Based on Geometric Discrete Mechanics"
Nov. 11, 2018
The following paper has been accepted.

T. Satoh and T. Yaguchi, On the equivalence of the norms of the discrete diffrential forms in discrete exterior calculus, Japan Journal of Industrial and Applied Mathematics.