Satoki Ishikawa

Matrix Norm Regularization for Collaborative Learning of Large-Scale Deep Learning Models

Grant No.：JPMJAX25C1

Researcher

Satoki Ishikawa

Outline

As the training cost of large-scale deep learning models grows exponentially, it has become increasingly impractical for a single research center to train such models from scratch. However, simply combining models independently trained at multiple sites often leads to performance degradation due to knowledge interference and catastrophic forgetting. In this study, we develop an efficient method for computing matrix-norm-based regularization, enabling the construction of a collaborative learning framework that allows multiple research centers to jointly train large-scale models.

Noboru Isobe

Pseudo-Euclidean Geometrization of Optimal Transport Theory for Sustainable AI Development

Grant No.：JPMJAX25C2

Researcher

Noboru Isobe

Outline

To make large-scale AI a sustainable technology, there is a pressing need to develop model compression methods with clearly defined mathematical principles. In this study, we focus on the differential equation model of Transformer—-a core technology of modern AI—-and aim to elucidate the relationship between the mathematical structure of the equations and the learning rules of the model, thereby developing predictively controllable compression techniques. The key to this technological development lies in constructing a novel optimal transport theory that enables the utilization of the underlying pseudo-Euclidean geometric structure embedded within the model.

Junsei Ito

Development of Sim2Real Transfer Methods under Partial Observability

Grant No.：JPMJAX25C3

Researcher

Junsei Ito

Outline

This research aims to develop Sim2Real transfer methods under partial observability constraints in real environments and to solve technical challenges in realizing Physical AI. Specifically, this study will focus on: (1) developing observation dimension-adaptive data generation methods, (2) constructing physics-consistency-preserving learning frameworks, and (3) introducing safety guarantee mechanisms based on control theory. Through the integration of these approaches, this research will demonstrate efficient knowledge transfer from simulation to real environments using autonomous mobile robots.

Michikuni Eguchi

Haptic generative AI with user-motion feedback

Grant No.：JPMJAX25C4

Researcher

Michikuni Eguchi

Outline

Enhancing user immersion in virtual experiences requires realistic haptic feedback that reflects the diverse components of body motion. In this work, I develop a real-time haptic generative AI that incorporates user-motion feedback by combining a large-scale haptic dataset, collected across diverse objects and motion conditions, and a low-latency generative AI architecture. This approach advances toward a general-purpose haptic engine capable of delivering realistic haptic stimuli across a wide range of virtual environments.

Takashi Otonari

Novel View Synthesis of Urban Scenes from Pedestrian-Captured Videos

Grant No.：JPMJAX25C5

Researcher

Takashi Otonari

Outline

This research aims to achieve high-fidelity novel view synthesis from casual videos captured by pedestrians. These videos in urban scenes pose significant challenges for 3D reconstruction, as they often lack sufficient parallax and scene information, and frequently contain dynamic objects. To overcome these limitations, my approach will leverage prior knowledge from large-scale data, including 3D foundation models and generative models. The ultimate goal of this work is to democratize 3D content creation, enabling anyone to easily create and share immersive 3D content.

Eri Onami

Foundational Approaches for Enhancing Legal Reasoning Capabilities of Large Language Models

Grant No.：JPMJAX25C6

Researcher

Eri Onami

Outline

In general, it is well established that large language models improve their performance by learning from massive datasets. On the other hand, in the legal domain, it is known that simple training on existing large-scale legal datasets alone is insufficient to achieve the level of performance required for effective AI utilization in legal practice. This study proposes the construction of training data with an emphasis on legal semantics, along with corresponding model training.

Kentaro Onda

Development of a speaker-invariant speech representation learning model based on structural phonology

Grant No.：JPMJAX25C7

Researcher

Kentaro Onda

Outline

According to structural phonology, which regards phonemes not as physical features but as representations based on their interrelations, the reason why we humans can understand speech regardless of the speaker is that we perceive an utterance as an overall structure. In this study, we revisit “structural representations” of speech, the technical implementation of structural phonology, and attempt to build a self-supervised learning model for speech that can extract speaker-invariant features. This research is expected to contribute to realizing spoken language processing that is closer to human perception.

Tonan Kamata

Toward a Unified Computational Framework Linking the Continuous and the Discrete with Real-World Challenges

Grant No.：JPMJAX25C8

Researcher

Tonan Kamata

Outline

This project models the complexity of real-world tasks with continuous structures—such as design, logistics, and control—through puzzle-like abstractions including dissection problems, the art gallery problem, and anti-slide puzzles. These are then formalized as decision problems over real variables. By proving membership and completeness results within complexity classes such as ER and EAR, we aim to clarify the theoretical boundary between tractable and intractable problems in real-world applications. In parallel, we pursue the development of general-purpose software leveraging symbolic computation libraries, thereby providing both conceptual insights and practical computational tools.

Yukino Kawai

Mathematical Information Foundation for AutoML Clustering

Grant No.：JPMJAX25C9

Researcher

Yukino Kawai

Outline

Although visual evaluation is the mainstream approach for clustering evaluation, quantitative automated evaluation techniques are essential for analyzing large-scale data. This research project involves developing metrics for cluster separation and cohesion that reflect graph structures, as well as efficient computational methods, inspired by combinatorial concepts that integrate splits and diversity from mathematical biology. These indicators will be applied to variable selection and multimodal data integration with the aim of contributing to the development of the next generation of AI through the deployment of AutoML clustering.

Yuri Kinoshita

Research on fundamental principles and development of models with interaction between memory and learning

Grant No.：JPMJAX25CA

Researcher

Yuri Kinoshita

Outline

While memory-augmented models have recently attracted much attention in artificial intelligence, theoretical foundations and practical implementations of an interactive system between memory and learning remain areas of active research. In this project, we analyze such interaction based on optimization theory to elucidate its fundamental principles. At the same time, we enhance its efficiency and autonomy employing these insights.

Hiroki Kusuyama

Pro×Cam×AI Collaboration for High-Speed Dynamic Projection Mapping towards Perceptual Realism

Grant No.：JPMJAX25CB

Researcher

Hiroki Kusuyama

Outline

This research aims to seamlessly integrate the real and virtual worlds in daily life by efficiently controlling a compact projector-camera system using AI technology. The proposed AI architecture enables high-speed dynamic projection mapping (DPM) without the need for markers on the projection target or cumbersome pre-calibration procedures. Additionally, by integrating the DPM system with a projection-based illumination system capable of high-contrast projection mapping even under bright environmental lighting, we seek to achieve perceptual realism, making the projected patterns appear as if they are inherent textures of the physical objects themselves.

Tatsuya Kubo

Large Reservoir Computing Using Commercial DRAM

Grant No.：JPMJAX25CC

Researcher

Tatsuya Kubo

Outline

As transistor scaling reaches its physical limits, a new design paradigm, Analog AI, has emerged that actively exploits process variation to enable faster and more energy-efficient AI computation. However, existing analog AI implementations remain limited to thousands of parameters, leaving their potential for modern AI applications largely unexplored. To bridge this gap, this research leverages the non-deterministic behavior of commercial DRAM when operated outside specifications to realize physical reservoir computing. By repurposing existing memory hardware into an analog AI substrate rather than fabricating specialized chips, this research aims to demonstrate the first billion-parameter analog AI system.

Mayuko Kori

Towards a unified theory and new methods for Bellman operators via category theory and fixed-point theory

Grant No.：JPMJAX25CD

Researcher

Mayuko Kori

Outline

This project aims to abstract Bellman operators from categorical perspective, and develop a unified framework instantiating various model-specific operators. This theoretically unifies learning methods traditionally designed heuristically. Furthermore, this project seeks to apply efficient fixed-point computation techniques developed in other fields, such as formal verification, in order to design learning methods that are not only theoretically sound but also effective even in settings where existing approaches face difficulties.

Machi Shimmei

Towards Effective Human-AI Interaction to Address Misconceptions in Statistics

Grant No.：JPMJAX25CE

Researcher

Machi Shimmei

Outline

Artificial intelligence (AI) is playing an increasingly important role in human decision-making processes. A key step toward building trustworthy AI involves understanding how to interact with individuals when they hold misconceptions. In this project, we investigate methodologies for identifying and addressing human misconceptions. To pursue this goal, we focus on the domain of statistics, develop a dataset of common misconceptions, and design an adaptive tutoring system that can diagnose students’ misunderstandings and provide tailored corrective feedback.

Wataru Setoyama

Development of Efficient Quantum Algorithm for Solving Nonlinear Differential Equations

Grant No.：JPMJAX25CF

Researcher

Wataru Setoyama

Outline

Quantum algorithms are expected to provide substantial computational speedups over classical methods, yet their effectiveness in solving nonlinear dynamical systems with important real-world applications remains limited. In this project, I extend quantum algorithms by exploiting a correspondence between quantum-mechanical representations and polynomial computation. The aim of this project is to identify a new class of nonlinear differential equations that can be solved on quantum computers and to develop systematic methods for addressing such problems.

Takehiro Takayanagi

Structural Analogy AI for Navigating Unseen Social Situations

Grant No.：JPMJAX25CG

Researcher

Takehiro Takayanagi

Outline

This study aims to develop a Structural Analogy AI that supports reasoning in highly uncertain domains such as society and the economy. While existing large language models excel at capturing semantic similarities, they remain limited in recognizing logical structural analogies. To address this gap, we construct reasoning graphs based on logical structures extracted from expert reasoning corpora. By enabling AI agents to explore these graphs, our approach derives reasoning paths with analogous structures, thereby facilitating hypothetical reasoning about previously unseen situations.

Shoji Toyota

Simulation-based Inference for Time Series Simulators

Grant No.：JPMJAX25CH

Researcher

Shoji Toyota

Outline

Simulation-based inference (SBI) is a machine learning method for identifying parameters of simulators from observed data. While many SBI approaches target time-series simulators, they often fail to fully exploit their temporal structure. In this project, we develop an efficient SBI framework that leverages the temporal structure of simulators. Furthermore, we investigate its connections with time-series statistical analysis, which performs statistical inference by utilizing temporal properties. Through this perspective, we attempt to integrate SBI with time-series statistical methods.

Riu Naito

Next-Gen Acceleration of Generative Models via Stochastic Calculus

Grant No.：JPMJAX25CI

Researcher

Riu Naito

Outline

The objective of this project is to construct a new algorithm for a next generation acceleration of generative diffusion models via stochastic calculus. In order to overcome the difficulty of a huge computational cost in the implementation of generative models, I employ a fast numerical algorithms for stochastic differential equations based on stochastic calculus. By making a breakthrough in accelerating generative models, this project aims to solve data generation challenges in medicine, engineering and economics where high-accuracy data are required.

Takumi Hara

Active Inference-Based Robotic Neck Control: “Motion Speaks Intent”

Grant No.：JPMJAX25CJ

Researcher

Takumi Hara

Outline

For autonomous robots to collaborate with humans, it is essential to possess the ability to naturally convey the “intent” behind their actions. In this study, based on the framework of active inference, I realize a control system that generates neck movements imitating gaze shifts according to the prediction error of spatial attention representing objects of interest in space. By correcting perception through action and “speaking” internal states through motion, my approach brings structural and implementational innovations to the explainability and social interaction capabilities that conventional AI has lacked.

Yasunari Hikima

Decision-Oriented Learning with Imperfect Observations

Grant No.：JPMJAX25CK

Researcher

Yasunari Hikima

Outline

Mathematical optimization has long played an important role in supporting complex decision-making by individuals and organizations. However, addressing uncertainty in the parameters of mathematical models remains a significant challenge. In this study, we pursue a decision-oriented learning approach to estimate the parameters of optimization models based on observed histories of past decisions. This study aims to advance data-driven decision support by complementarily integrating mathematical optimization with machine learning.

Yosuke Fukuchi

Cognitive AI Agent for Simulating Distorted Linguistic Activities in Online Media

Grant No.：JPMJAX25CL

Researcher

Yosuke Fukuchi

Outline

This study develops a cognitive AI agent that integrates cognitive science theories with large language models to reveal the thought processes of users who adopt false beliefs or extreme ideologies through web media. The agent aims to simulate distortions in language-based cognitive activities. In particular, this study seeks to reproduce in detail the subtleties of users’ every action, the shifts in their underlying thoughts, and their cognitive instability, thereby contributing to the design of healthier digital environments.

Satoki Masuda

Understanding and Controlling Dynamic Location Equilibrium through Time-Series Synthetic Data

Grant No.：JPMJAX25CM

Researcher

Satoki Masuda

Outline

To understand the dynamics of urban land use, it is essential to estimate the interactions among the decision-making of residents, government authorities, developers, and other stakeholders. In this study, we automatically generate time-series networks of land use and transportation from open data and integrate them with synthetic micro-level residential relocation data. By utilizing these multi-agent, time-series behavioral data, we estimate a dynamic location model that fuses data-driven approaches with equilibrium theory, enabling high-precision prediction of land use and population mobility. Furthermore, we develop an urban planning simulation platform based on the model.

Youmi Ma

Improving the Long-Context Processing Ability of Large Language Models via Self-Distillation

Grant No.：JPMJAX25CN

Researcher

Youmi Ma

Outline

Large Language Models (LLMs) exhibit a fundamental limitation that their performance tends to degrade as the length of context being processed increases. This study aims to overcome this limitation by enabling LLMs to acquire and enhance their long-context processing abilities through self-distillation, using their own intermediate representations and outputs from short-context inference as guidance. In this way, this study not only seeks an efficient way to build LLMs that can robustly handle long contexts, but also aims to elucidate the underlying mechanisms through which LLMs process contextual information.

Yukiya Mita

An interactive AI accompaniment system designed to achieve a deep understanding of a performer’s expressive intent and the underlying structure of a musical piece.

Grant No.：JPMJAX25CO

Researcher

Yukiya Mita

Outline

This year’s research aims to establish a foundational understanding of musical scores and a model for a performer’s expressive style. We will integrate implicit knowledge, such as music theory and historical context, into a Large Language Model (LLM) using Retrieval-Augmented Generation (RAG) technology, going beyond the symbols written on the score. Furthermore, we will quantitatively model the relationship between notational symbols and acoustic nuances by deeply understanding it through contrastive learning on a performer’s personal data. This will lay the groundwork for a future AI accompaniment system capable of multifaceted, semantic expression comprehension.

Yuri Morizane

Development of Inverse Design Platform for Neutron Spectrum

Grant No.：JPMJAX25CP

Researcher

Yuri Morizane

Outline

Neutrons have great potential for use in a variety of fields, including cancer therapy and nuclear fusion. However, achieving the ideal neutron spectrum for each application requires considerable time and effort. My aim is to construct an AI capable of inverse design of the material composition and geometric information of the moderator required to achieve the target neutron spectrum incorporating the relevant physical laws. This will allow us to develop a model based on both theory and experimentation.

Ryutaro Mori

Extending Mechanism Design with LLM Agents as Human Proxies

Grant No.：JPMJAX25CQ

Researcher

Ryutaro Mori

Outline

To resolve social dilemmas—conflicts between individual and collective interests found in areas like commons management or CO2 emission regulations—communication among stakeholders is crucial. However, the optimal mechanisms to facilitate such communication remain largely unknown. This project will develop AI agents that work as a proxy for human decision-making and utterances and use these agents in large-scale simulations to identify optimal communication mechanisms. A key aspect of this approach is explicitly accounting for the behavioral differences between AI agents and humans. This is essential for discovering robust solutions applicable in the real world and paving the way for a new methodology to design better social systems.

Satoshi Yamamori

Discovering Embodiment Hierarchies through Representation Learning

Grant No.：JPMJAX25CR

Researcher

Satoshi Yamamori

Outline

Dexterous whole-body control of heavy objects remains a challenging task, highlighting the potential for integrating artificial intelligence and robotics. This study focuses on the spatial and temporal hierarchies inherent in dynamic tasks and embodied systems, aiming to achieve sensory processing and motor control across multiple time scales and modalities. To this end, we develop a framework based on latent representations of dynamical systems that enables the acquisition and communication of hierarchical structures. Through this approach, we construct an innovative learning theory that realizes human-like motion generation in real-world robotic systems.

Ryo Yoshida

Interpretable and High-Performance LLMs Grounded in Cognitive and Neural Data

Grant No.：JPMJAX25CS

Researcher

Ryo Yoshida

Outline

This project aims to establish a methodology for fine-tuning LLMs to align their internal mechanisms with human cognitive and neural data, achieving both high interpretability and performance. Specifically, I will: (i) establish an information-theoretic framework for aligning LLM internal mechanisms with cognitive and neural data, (ii) interpret LLM internal mechanisms through the lens of cognitive and neural data, and (iii) fine-tune LLM internal mechanisms based on cognitive and neural data.

Hirotaka Yoneda

Improving graph coloring algorithms

Grant No.：JPMJAX25CT

Researcher

Hirotaka Yoneda

Outline

We study approximation algorithms for graph coloring, a fundamental problem in theoretical computer science. In 2024, Kawarabayashi, Thorup, and Yoneda proposed an algorithm to color 3-colorable graphs with O(n^0.1975) colors. We aim to improve the results further by discovering new local and global characteristics of 3-colorable graphs.

Kento Watanabe

AI Systems Bridging Music Information and Real-World Human Motion

Grant No.：JPMJAX25CU

Researcher

Kento Watanabe

Outline

Recent advances in deep learning have boosted music generation and analysis, yet a framework that jointly handles music information (audio and lyrics) and real-world human motion (dance, singing, instrumental performance) remains underdeveloped. We aim to quantify when and how rhythm, melody, and lyrical prosody align with limbs, fingertips, and facial expressions, and to provide an integrated model usable for creation and education. This project performs unified analysis of audio/lyrics and human motion, and builds an AI platform that visualizes, searches, and generates these cross-modal correspondences.