Research
Research
We aim to understand and advance modern machine learning models from a human-centered perspective, while simultaneously using machine learning models to investigate human cognition and brain activity. Here, we introduce representative research projects currently underway.
Research Topics
🚧 Under construction 🚧
Development of Large Language Models and Exploration of Internal Representations
As part of an interdisciplinary collaborative organization involving over 1,500 participants, we developed and publicly released a high-performance Japanese-language large language model as a fully open resource. Leveraging this open LLM development, we comprehensively investigate internal representations, learning dynamics, and instruction optimization, providing integrated insights that balance improved model performance with enhanced interpretability.
Generative AI and Brain Activity
We pursue research that integrates state-of-the-art generative models with brain activity measurements, aiming to bridge information perceived and recalled by the human brain with information represented internally within generative AI systems. In this project, we construct a general-purpose framework applicable to both visual and auditory modalities, and systematically examine the correspondence principles between internal representations in generative AI and neural representations in the brain. Through collaborations with industry partners such as Google Research, we accelerate both the practical applications of multimodal brain decoding technologies and the mutual understanding of information processing mechanisms in machines and humans.
Decision-Making and Brain Activity
We aim to globally and temporally resolve how distributed signals within the brain are integrated and culminate in rapid decision-making. By combining millisecond-scale whole-brain activity obtained through non-invasive measurements with mathematical methods that extract shared computations across brain regions, we seek to reconstruct the full picture of “distributed decision-making circuits” that are not confined to individual neurons or single brain areas.
Selected Research Outputs
For a more detailed list of publications, please refer to Google Scholar and Researchmap.
Correspondence Between Music Generative AI and Human Brain Activity (Nature Communications 2025)
We successfully aligned music generative models with fMRI brain activity and reconstructed music directly from neural signals. The generated music preserved semantic attributes such as genre and mood, demonstrating that representations in music AI correspond to brain activity in and around the auditory cortex. This work was conducted in collaboration with Google DeepMind and was featured in Google Research’s official YouTube series, Field Notes.
Neural Basis of Cross-Modal Representations and Emotion (Nature Communications 2025, Editor’s Highlights)
Using feature representations from dance generative models, we analyzed brain activity during natural dance observation. We found that cross-modal representations explain neural responses better than low-level motion or acoustic features. Furthermore, we quantitatively characterized differences in neural representations associated with emotion and expertise. This paper was selected as an Editor’s Highlight by Nature Communications.
Learning Principles of Supervised Fine-Tuning in Large Language Models (EMNLP 2025)
We systematically analyzed how supervised fine-tuning (SFT) alters large language models. By comparing more than 1,000 SFT-trained models, we identified learning effects that are shared across tasks as well as model-dependent differences.
Correspondence Between Large Language Models and Human Brain Activity (EMNLP 2024)
We analyzed brain activity recorded during approximately eight hours of movie and TV drama viewing, using large language models to characterize spoken content, situational context, and narrative structure, and revealed their corresponding neural representations and distributions.
Relationship Between Image Generative AI and Human Brain Activity (CVPR 2023)
We demonstrated a quantitative relationship between brain activity and latent representations in image generative AI, enabling the visualization of perceptual content directly from neural signals. This work was covered by more than 60 media outlets, including Science, Newsweek, and Asahi Shimbun.
Visualization of Neural Dynamics of Decision-Making Using Non-Invasive Brain Measurements (eLife 2021)
By decomposing non-invasive MEG signals along multiple task dimensions, we reconstructed the trajectory of decision-making processes from sensory input to motor response with millisecond temporal precision.
A Method for Extracting Shared Computational Information Across Multiple Brain Regions (NeurIPS 2020, Spotlight)
We developed a novel method (dSCA) for analyzing simultaneous recordings across multiple brain regions. The method separates task-related variables while extracting computational components shared across regions, and visualizes when and which variables are shared over time, providing new insights into multi-region cooperative computations underlying decision-making.
and more…