EVENTイベント情報
We are pleased to announce that “AIE & GP-DS International Symposium” is going to be held.
Program
| Chair Prof. Yoshizawa |
9:00-9:10 | Openning remarksProf. Kazama |
|---|---|---|
| 9:10-9:25 | Introduction of AIE ProgramProf. Kaneko | |
| 9:25-9:40 | Introduction of GP-DS ProgramProf. Nakao | |
| 9:40-9:45 | Announcement from Program chairAIE program chair | |
| Chair Prof. Nakao |
9:45-10:25 |
Keynote speech 1
Semantic Science: Ontology, Graph Neural Networks and Knowledge Graphs for FAIR, Learning, and ReasoningProf. Roger H. FrenchProfessor, Materials Data Science for Stockpile Stewardship Center of Excellence, Case Western Reserve University |
| 10:25-10:30 |
Break | |
| 10:30-12:30 |
1 minute oral poster summary & Poster presentationsGP-DS & AIE students | |
| 12:30-14:00 | Lunch break | |
| Chair AIE program chair |
14:00-16:00 |
1 minute oral poster summary & Poster presentationsAIE & GP-DS students |
| 16:00-16:10 | Break | |
| Chair Prof. Sakaguchi |
16:10-16:50 |
Keynote speech 2
Natural Language-based Assessment and Feedback for Second Language Learners of EnglishProf. Kate KnillResearch Professor, Cambridge University |
| 16:55-17:35 |
Keynote speech 3
Open and Repurposable: Foundation Models for the Automotive IndustryDr. Andrei BursucDeputy Scientific Director at valeo.ai | |
| 17:35-17:40 | Closing and best poster presentation awardsGP-DS program chair | |
| chair AIE |
17:50-19:30 | ReceptionAIE & GP-DS |
Keynote speech 1
Semantic Science: Ontology, Graph Neural Networks and Knowledge Graphs for FAIR, Learning, and Reasoning
Prof. Roger H. FrenchProfessor, Materials Data Science for Stockpile Stewardship Center of Excellence, Case Western Reserve University
Abstract
Machine learning and reasoning on sparse, massive, and historical science datasets to address novel and long-standing materials science problems is becoming possible with advances in the Semantic Web and Materials Data Science (MDS). The foundations of Materials Data Science (MDS) computing, include 1) the integration of distributed (e.g., Cloudera, Hadoop) and high performance computing, 2) spatiotemporal Graph (st-Graph) Neural Network Foundation models, 3) Resource Description Framework (RDF) triplestore Graph Databases (e.g., Ontotext’s GraphDB) and 4) automated analysis pipelines and scientific workflows. Integration of these with Study Protocols are an essential feature for robust, data-centric, MDS studies, where data FAIRification is enabled by the MDS-Onto1 ontology and tools, transforming messy data into RDF linked data2,3 across Study Stages, including sample, tool, recipe, pre-processing, analysis, modeling and results publishing. These concepts, relationships and stages guide the research process allowing exchange of serialized JSON-LD files and their deserialization with Apache Arrow into language-independent dataframes for analysis, modeling, and reasoning. This Materials Data Science framework, provides the ability to curate AI-ready data and knowledge for the era of Materials Science Foundation Models.
[1] https://cwrusdle.bitbucket.io/
Biography
Roger H. French is the Kyocera Professor in the Case School of Engineering and a Distinguished University Professor, Case Western Reserve University, Cleveland, Ohio. His primary appointment is in Materials Science and Engineering, with secondary appointments in Computer and Data Sciences,Macromolecular Science, Biomedical Engineering, and Physics.
He is the director of the DOE-NNSA Center of Excellence for Materials Data Science for Stockpile Stewardship. He is the faculty director of the CWRU Applied Data Science program which offers graduate courses and graduate certificates and an undergraduate minor university-wide. He is the director of the SDLE Research Center, an Ohio Third Frontier center focused on advancing materials data science, big data analytics, and reliability of long-lived technologies.
Before this he was in Central Research at the DuPont Co. in Wilmington, Delaware since 1985 and was Adjunct Prof. of Materials Science at the Univ. of Penn. since 1996.
Keynote speech 2
Natural Language-based Assessment and Feedback for Second Language Learners of English
Prof. Kate KnillResearch Professor, Cambridge University
Abstract
The rapid rise of large language models (LLMs) is reshaping computer-assisted language learning and automated assessment of English second language (L2) learners. Our proposed Natural Language-based Assessment (NLA) framework leverages LLMs to interpret and apply established natural language scoring criteria and descriptors related to specific aspects of proficiency – such as those from the Common European Framework of Reference (CEFR) –traditionally used by human examiners. By “connecting the dots” between these can-do descriptors and learner productions, NLA enables nuanced, human-aligned assessment and feedback in a scalable, automated manner which supports language learners, teachers, and testers. Our approach targets individual dimensions of proficiency – such as fluency, coherence and cohesion, grammatical accuracy – for both holistic and analytic assessment. We also apply NLA to generate fine-grained, constructive natural language feedback, highlighting what learners can do, not just what they cannot. By grounding assessments in established and interpretable descriptors, the approach enhances the transparency and explainability of LLM outputs addressing a key limitation of black-box methods commonly found in automated, deep learning-based assessment. Moreover, NLA can be applied in a zero-shot setting, requiring no training data, improving accessibility, reducing costs, and avoiding biases introduced through fine-tuning.
Biography
Prof. Kate Knill is a Research Professor at the Department of Engineering and the Automatic Language Teaching and Assessment Institute (ALTA), Cambridge University. She is the Principal Investigator for ALTA's Spoken Language Processing (SLP) Research. She has over 30 years’ experience developing automatic speech recognition, text-to-speech synthesis and auto-marking systems in industry and academia. As an individual researcher and a leader of multi-disciplinary teams as Languages Manager, Nuance Communications, and Assistant Managing Director, Toshiba Research Europe Ltd, Cambridge Research Lab, she has developed speech systems for over 50 languages and dialects. Her current research focus is on applications for non-native spoken English language assessment and learning, and detection of speech and language disorders. She is a Senior Member IEEE and a Fellow of the International Speech Communication Association (ISCA). She is Chair of IEEE Recognitions Council. She was a member of the IEEE Speech and Language Processing Technical Committee (2009-2012) and the IEEE James L. Flanagan Speech and Audio Processing Award Committee (2020-25, Chair 2023-24), ISCA Board member 2013-2021 and ISCA Secretary 2017-2021. She received BEng from University of Nottingham and PhD degrees from Imperial College, University of London, respectively, both sponsored by Marconi Underwater Systems Ltd.
Keynote speech 3
Open and Repurposable: Foundation Models for the Automotive Industry
Dr. Andrei BursucDeputy Scientific Director at valeo.ai
Abstract
Real-world perception for autonomous driving (AD) requires models that can learn from massive, uncurated, multi-sensor datasets with minimal, costly labeling. Self-supervised learning (SSL) is key to training foundation models in this regime. This talk moves beyond canonical SSL to present open-source foundation models from the valeo.ai team. We will highlight key examples, including Franca, our fully open-source foundation vision encoder (data, models, and code), VaVIM/VaVAM, our open generative video and video-action models, and DrivoR, our state-of-the-art camera-based end-to-end autonomous driving model. We focus on how these models can be built on top of each other, creating a composable stack for perception. We'll show how this enables advanced applications like cross-sensor distillation, auto-labeling, and architecture repurposing, ultimately saving time, money, and improving performance.
This work exemplifies why open models are critical for developing the next generation of annotation-efficient and reliable autonomous systems.
Biography
Andrei Bursuc is a senior research scientist and deputy scientific director at valeo.ai and research associate at the Astra Inria project team in Paris working on perception for assisted and autonomous driving. Previously he was a research scientist at Safran Tech in the aerospace-defense industry. Prior to that he was a postdoctoral researcher at Inria Paris, within the Willow project team working with Josef Sivic and Ivan Laptev, and Inria Rennes with Hervé Jégou. He did his PhD at Ecole des Mines Paris and Alcatel-Lucent Bell Labs France with Francoise Preteux and Titus Zaharia on visual content indexing and retrieval.
His research interests concern reliability of deep neural networks, learning with limited supervision (self-supervised, unsupervised, few-shot learning) and multi-modal multi-sensor perception. On reliability he is focused on pragmatic uncertainty estimation inspired from Bayesian approaches, e.g., PackedEnsembles, Bayesian Neural Networks from a single training run. On self-supervised learning he is currently focused on foundation models for other modalities and sensors (Lidar, multi-camera rigs) and strategies to efficiently adapt them to other data distributions (from internet content to urban driving).
Andrei has also been teaching at Ecole Polytechnique the “Deep Learning Do It Yourself!” course with Marc Lelarge and is a member of the ELLIS program. Andrei is regularly part of the technical program committee for CVPR, ICCV, ECCV and NeurIPS, but also co-organized multiple tutorials and workshops at the same venues in the past years.
Poster session 1: GP-DS ＆ AIE
GP-DS
Computing for human well-being: from physiological sensing to psychological reflection
Abstract
Human-Computer Interaction (HCI) area offers opportunities to address both the physiological and psychological challenges connected with digital life. This poster shows two distinct approaches to enhancing user well-being. First, addressing physical ergonomics in virtual environments, we introduce EyeXRciser, a gaze redirection method designed to mitigate eye strain during VR reading. By passively shifting the text window within a head-bound coordinate system, the system induces eye movements to prevent muscle stiffness. A user study (N=24) demonstrates that using an unnoticeable redirection speed (0.03 rad/s) significantly minimizes the decline in accommodative ability without causing discomfort or impairing reading comprehension. Complementing this physiological intervention, our second work targets mental well-being from motion prediction perspective. We introduce a privacy-preserving framework integrating mmWave radar with Multimodal LLMs to overcome the subjectivity and privacy issues of traditional monitoring. This system extracts objective behavioral insights from daily activities, enabling non-intrusive, longitudinal assessment of diverse mental health conditions.
Tracking-by-Detection in Multi-Target Multi-Camera Tracking
Abstract
Tracking-by-detection has become a dominant paradigm in single-camera multi-object tracking (MOT) due to its simplicity and effectiveness. Although tracking-by-detection is widely adopted in single-camera MOT, it does not directly extend to multi-camera tracking due to fragmented cross-view detections. To extend tracking-by-detection to online multi-target multi-camera tracking (MTMCT), we reformulate the Detection stage by constructing multicamera detections as clusters of single-camera detections, enabling MTMCT to be decomposed into sequential stages of Multi-Camera Detection, Association, and Re-Identification.
Balancing energy and computation: metabolic strategies of the brain across behavioral states
Abstract
Energy is fundamental for driving information processing. In the brain, each action of potential firing and synaptic plasticity induction consumes energy. Consequently, ATP production, energy substrate availability, and metabolic exchange among blood vessels, astrocytes, and neurons constrain how neuronal circuits encode information and store memories. However, how metabolic supply adapts to shifting computational demands across brain states remains poorly understood. Here, we investigated how vasculature and astrocytes respond to the energy demands of rapid eye movement (REM) sleep. REM sleep is hypothesized to optimize stored experiences, as reflected in human dreams, and is considered highly energy demanding. To examine metabolic adaptation, we used wide-field fluorescence imaging in mice expressing a fluorescent albumin-mScarlet conjugate in blood plasma, a FRET-based pyruvate sensor in astrocytes, and an ATP sensor in neurons. We found that energy dynamics are strongly coupled to neuronal activity and that the spatiotemporal patterns of blood volume fluctuations are sleep-state-dependent. Surprisingly, during REM sleep, we observe a marked increase in blood volume and astrocytic pyruvate, yet a paradoxical drop in neuronal ATP levels. These findings reveal a dynamic, state-dependent interplay between neurons, astrocytes, and vasculature, supporting the brain’s energy efficiency and flexible information processing across sleep-wake cycles.
Turbulence modeling using a constrained artificial neural network
Abstract
Turbulent flow is a type of flow where the fluid undergoes chaotic changes in its velocity and pressure. It is important to understand the characteristics of turbulent flow because most of the flow we see in our environment is likely to be in a turbulent state. However, the irregular behavior of turbulent flow is often difficult to predict even with modern supercomputers because resolving all the scales, from large eddies to small ones, is computationally intensive. One of the methods to alleviate this problem is to calculate only the large-scale flow, filtering out the small-scale flow. However, it remains impossible to derive the exact governing equations for this method, which is known as the closure problem. Therefore, it is common to employ a model usually based on physical insights, introducing some empirical parameters. To improve the model’s performance, it is desirable to derive the functional relationships between the large-scale flow and the small-scale flow with fewer assumptions. Therefore, here we use an artificial neural network to infer the functional relationship without any explicit assumption. In addition, we improve the model’s performance by adding a constraint to ensure that the prediction is physically appropriate.
Cohort Study on the Validity and Reliability of the Scale for Suspicion of Psychogenic Non-Epileptic Seizures
Abstract
Diagnosing psychogenic non-epileptic seizures (PNES) is clinically fundamental but challenging. Video-electroencephalography monitoring (VEEG) is ideal but not always applicable to all patients. The scale for suspicion of PNES (SS-PNES), a simple and practical tool, has demonstrated reliability in Brazil. This study aims to validate the SS-PNES in Japan. We evaluated 250 patients who underwent comprehensive epilepsy evaluation, including VEEG, neuropsychological and psychosocial assessment at Tohoku University Hospital from January 2020 to April 2023. The SS-PNES comprises 8 items assessing epileptological aspects and 7 items related to current and past psychiatric/psychological characteristics, using 4-point Likert scale. Exploratory factor analysis (EFA) indicated that the essential structure for diagnosing PNES required only 6 of the 15 items, with two factors related to emotional stress and psychological disorders. We demonstrated the superiority of the two-factor model and supported its construct, factorial, and convergent/divergent validity. Cronbach’s alpha coefficient was 0.66 and 0.60 in Factor 1 and 2. The area under the ROC curve was 0.89. The SS-PNES can be used to suspect PNES even with fewer psychiatric/psychological characteristics. Additional epileptological information may be necessary to diagnose the presence of epilepsy, either with or without PNES, which could enhance its application in future Japanese epilepsy practice.
AIE
Manzai Analysis: The Influence of Proficiency on Tempo
Abstract
This research aims to clarify the factors that determine the “funniness” of Manzai.
As a first step, we investigate the influence of proficiency on tempo. In Manzai, tempo is a key element in eliciting laughter. Controlling tempo is a sophisticated skill, and like many other complex skills, experience is considered essential for mastery. However, in the context of Manzai, it remains unclear specifically how this difference in experience is reflected in tempo. To address this, we focus on the “switching pause,” a component of tempo defined as the time interval until the next speaker begins. Predicting that switching pause tendencies vary by pair, we treated them as features of each pair. We analyzed the switching pauses of skilled performers (M-1 Grand Prix finalists) and novices (university students with one year of experience) and confirmed the possibility that there are differences in tendencies between the two groups. In future work, we plan to expand the dataset and further investigate the differences in switching pause tendencies.
Reconfiguring Independent Sets and Vertex Covers Under Extended Reconfiguration Rules
Abstract
In reconfiguration problems, we are given two feasible solutions to a graph problem and asked whether one can be transformed into the other via a sequence of feasible intermediate solutions under a given reconfiguration rule. While earlier work focused on modifying a single element at a time, recent studies have started examining how different rules impact computational complexity. Motivated by recent progress, we study Independent Set Reconfiguration (ISR) and Vertex Cover Reconfiguration (VCR), which are fundamental problems in this field, under the k-Token Jumping (k-TJ) model. In k-TJ, up to k vertices may be replaced from a current solution to a next solution. In this work, we show that the computational complexity of ISR and VCR crucially depends on the value of k. Furthermore, we clarify the difference in computational complexity between these problems when k, given as part of the input, is always slightly smaller than the solution size, whereas the problems themselves are equivalent when k is constant.
A Seat Surface Device with Dynamically Adjustable Softness
Abstract
Chairs used in everyday life employ seat surfaces with different levels of softness depending on their intended purposes and contexts; however, a single chair cannot present multiple levels of softness. In this study, we propose a seat-surface device that can be attached to a variety of chairs and dynamically control softness. The proposed system leverages the jamming transition of granular materials by regulating the internal air pressure of a sealed bag filled with polystyrene beads through controlled air inflow and outflow using a vacuum pump.
In this poster session, we present the design and implementation of the seat-surface device based on granular jamming and report a technical evaluation focusing on the achievable range of softness and its response characteristics, demonstrating that continuous control of softness is feasible. We further discuss potential applications of the system, including supporting desk workers’ concentration or relaxation depending on work context, gently regulating dwell time in open spaces, and balancing initial sitting comfort with ease of standing up.
A Mechanism of knowledge recognition in language models.
Abstract
To build language models (LMs) that are faithful to their own knowledge, it is important to understand the mechanisms by which they determine whether an input knowledge statement is something they can internally recall—that is, whether the knowledge is known or unknown.
In this study, we examine the hypothesis that an LM judges the existence of own knowledge based on the consistency between the words it internally recalls in response to an input and the actual input itself.
Through intervention experiments that manipulate knowledge recall, our results suggest that this consistency between prediction and input contributes to the LM’s knowledge recognition.
These findings provide insights into understanding the introspective behavior of language models.
Are Checklists Really Useful for Automatic Evaluation of Generative Tasks?
Abstract
Automatic evaluation of generative tasks using large language models faces challenges due to ambiguous criteria. Although automatic checklist generation is a potentially promising approach, its usefulness remains underexplored. We investigate whether checklists should be used for all questions or selectively, generate them using six methods, evaluate their effectiveness across eight model sizes, and identify checklist items that correlate with human evaluations. Through experiments on pairwise comparison and direct scoring tasks, we find that selective checklist use tends to improve evaluation performance in pairwise settings, while its benefits are less consistent in direct scoring. Our analysis also shows that even checklist items with low correlation to human scores often reflect human-written criteria, indicating potential inconsistencies in human evaluation. These findings highlight the need to more clearly define objective evaluation criteria to guide both human and automatic evaluations.
Exploring Social Norm Formation Through Large Language Model–Based Simulations
Abstract
This study investigates the potential of large language models (LLMs) for conducting social simulations to reproduce and examine the processes and mechanisms underlying the formation, maintenance, and transformation of social norms. It proposes a conceptual framework to guide future research in this area. By leveraging the ability of LLMs to emulate human-like behaviors—such as memory, decision-making, and action—we construct LLM-based agents capable of mimicking human conduct (LLM agents). Using these agents, we simulate interactions across different scenarios and analyze how social norms emerge, stabilize, and change through agent-to-agent dynamics.
Integrated Model of Ordinary Differential Equations and a Hidden Markov Model, and Its Application to Real-World Data
Abstract
Social phenomena cannot always be explained by a single equation.
To address this limitation, we constructed a mathematical model that integrates a hidden Markov model with ordinary differential equations, allowing a single social phenomenon to be described by multiple governing equations. This integrated framework enables the identification of latent social structures underlying observable social dynamics. We applied this model to analyze temporal changes in the proportion of social welfare recipients.
The results indicate that the observed dynamics can be explained by regime switching between two differential equations. By explicitly modeling both the underlying social structure and the data-generating mechanisms through differential equations, our approach provides insights that cannot be obtained using simpler methods, such as regression analysis, which primarily elucidate associations between variables.
A Survey on Algebraic and Transcendental Numbers and Their Variants
Abstract
The study of algebraic and transcendental numbers has been a central theme in number theory for a very long time and has led to deep interactions. In 2020, J. Rosen introduced finite algebraic numbers, which can be regarded as a finite analogue of classical algebraic numbers. Remarkably, the set of finite algebraic numbers is Q-algebra, closely paralleling that of the field of algebraic numbers. Associated with finite algebraic numbers, one can define certain periods, which exhibit nice properties analogous to those of classical complex periods. In this talk, we present a survey of the research conducted so far on finite algebraic numbers and their periods, focusing on their basic definitions, algebraic structures, and known results.
Poster session 2: AIE & GP-DS
AIE
Numerical Study of Multicomponent Fractional Quantum Hall Systems
Abstract
Two-dimensional electron systems can be realized at the interfaces of semiconductor heterostructures such as GaAs/AlGaAs, as well as in two-dimensional materials such as graphene. At low temperatures and under strong magnetic fields, these systems exhibit the quantum Hall effect, characterized by a vanishing longitudinal resistance and a quantized Hall resistivity. When the Landau level filling factor is fractional, the system is known as fractional quantum Hall system, where the electronic state is a strongly correlated quantum many-body state in which electron–electron interactions play a central role. Depending on the filling factor, fractional quantum Hall systems can host a variety of electronic states. While most previous studies have focused primarily on single-component systems, multicomponent systems, such as spinful electron systems and bilayer systems, are expected to support a much richer set of electronic states, including states relevant to topological quantum computation. Motivated by this background, I performed numerical calculations to investigate the electronic states realized in such multicomponent fractional quantum Hall systems.
Biological tissue analysis by mid-infrared photoacoustic spectroscopy based on ultrasound detection
Abstract
Mid-infrared photoacoustic spectroscopy allows for the non-invasive and molecularly selective analysis of biological tissues. We have developed a system that uses mid-infrared pulses with a repetition rate in the hundreds of kilohertz to induce photoacoustic signals, which are then detected using a piezoelectric transducer. While previous studies have reported a nonlinear relationship between photoacoustic signal amplitude and optical absorption coefficient, the underlying physical mechanism has remained unclear. In this study, we performed finite element method (FEM) simulations and systematically investigated the dependence of the photoacoustic response on optical absorption in a biological tissue model. This approach provided us with valuable insights into the origin of this nonlinearity.
Building on this understanding, we explored a resonance-based photoacoustic spectroscopy approach involving modulation of the pulse frequency to enhance signal generation. Using this method, we acquired photoacoustic spectra of human skin and performed measurements before and after meals over multiple days, in order to evaluate the feasibility of non-invasively analyzing blood-related components.
On Battery-aware Offloading for Renewable Energy Utilization in Virtualized RAN
Abstract
In 5G and beyond, virtualized radio access networks (vRANs), in which base station (BS) functions are virtualized and hosted on general-purpose servers, are expected to be introduced. By leveraging virtualization technologies, vRANs enable flexible computational resource allocation among BSs. As RANs become more advanced, however, their energy consumption continues to increase. To reduce energy costs and greenhouse gas emissions, we focus on integrating renewable energy sources, such as batteries and photovoltaic panels, into distributed BSs in vRANs. However, the variability of renewable energy generation causes spatial and temporal imbalances between energy supply and demand, resulting in battery overflows or the use of non-renewable energy. To tackle this problem, this study aims to improve renewable energy efficiency by offloading processing tasks among BSs according to energy supply–demand conditions. Specifically, we propose an offloading method based on the battery energy state.
A Study on Depth Accuracy Improvement of Indirect Time-of-Flight Image Sensor with In-Pixel Memory Array
Abstract
Automation of tasks and mobility systems has attracted considerable attention as a solution to labor shortages and as a means of enabling new transportation modalities. Safe and reliable automation requires Time-of-Flight (ToF) image sensors capable of understanding complex three-dimensional spatial structures. However, conventional ToF image sensors suffer from an inherent trade-off between frame rate and depth noise. In this study, to achieve safer and more reliable automation, we aim to improve this trade-off by introducing a three-dimensional stacked architecture, a backside-illuminated structure, and a pseudo multi-tap configuration. Through the adoption of these structural technologies, the proposed sensor achieves a reduced sensor size while securing more than three times larger aperture area and twice the in-pixel memory capacity compared to conventional sensors. As a result, theoretical calculations demonstrate that the proposed sensor can achieve approximately three times higher frame rate and more than 60% reduction in depth noise. This work is expected not only to contribute to the advancement of an automated society but also to promote progress in mobile application fields such as AR/VR, where precise and delicate gesture interaction is required.
High-Resolution, High-Speed Soft X-ray CMOS Image Sensors for Quantum Radiation with Low-Data-Rate Readout Technology
Abstract
This study develops a high-performance CMOS image sensor for soft X-ray detection, targeting applications such as synchrotron radiation facilities, semiconductor EUV inspection, and electron beam detection. Soft X-rays, with energies from 100 eV to 1 keV, are highly sensitive to electronic structures of materials, requiring detectors with high spatial resolution, wide dynamic range (WDR), high frame rate, global shutter operation, and strong radiation tolerance. To address challenges in conventional sensors—namely large pixel size and increased output data rate—this work introduces a novel in-pixel light intensity determination circuit that reduces output data by selectively reading signals according to incident intensity. A two-stage LOFIC architecture enables WDR operation, while analog memory-based voltage-domain global shuttering reduces circuit complexity. Furthermore, three-dimensional stacking technology separates the photodetector and readout circuits, significantly shrinking pixel pitch and increasing pixel count. The proposed sensor achieves improved dynamic range, higher saturation capacity, and high-speed readout. In addition, this study demonstrates the feasibility of applying soft X-ray sensor technology to electron beam detectors, highlighting its versatility for quantum beam applications and broader CMOS image sensor technologies.
Functional response and self-repair processes after focal injury in artificial neuronal networks with hierarchically modular structure.
Abstract
The brain, a biological information-processing system, can recover function after injury through flexible reorganization. Studies using animal and computational models have shown that functional recovery depends not only on lesion size but also on lesion location. However, because the brain is large-scale and highly complex, the mechanisms underlying injury-induced functional alteration and subsequent recovery remain incompletely understood.
In this study, we constructed an experimental platform based on a laser microdissection approach that enables selective damage to specific connection sites in an artificial neuronal circuit reproducing the hierarchical modular architecture observed in the mammalian cerebral cortex.
A microfluidic device was designed to implement a hierarchical modular structure consisting of sixteen 100-µm-square modules and was fabricated by replica molding using polydimethylsiloxane (PDMS). On days 10–11 in vitro, inter-module neurites were selectively severed using a UV nanosecond pulsed laser.
Immediately after lesioning, spontaneous activity was almost completely abolished in the vicinity of the damaged site, whereas activity was preserved outside the lesioned region, although its frequency was slightly reduced. In addition, clear differences in post-lesion recovery speed and activity changes were observed between hub connections located near the network center and peripheral connections located at the network edge.
Practical Optimization Method for Large-Scale Flux-Modulated-Type Magnetic Gears
Abstract
Magnetic gears, which can transmit power without any mechanical contact, offer lower vibration and acoustic noise compared to mechanical gears. Among the various types of magnetic gears, the flux-modulated-type magnetic gear has recently garnered attention because of its higher torque density and efficiency. It consists of concentric inner and outer rotors with pole pieces placed between the two rotors. The finite element method (FEM) is widely used for the design and analysis of electric machines, and using a partial model based on the periodicity of the electromagnetic field is an effective way to reduce model size and calculation time. However, for flux-modulated-type magnetic gears, their partial models cannot be minimized due to the low periodicity of the electromagnetic field. As a result, the model size remains large, and calculation time is still long. I propose a practical design method for flux-modulated-type magnetic gears by creating a partial model that focuses only on the inner magnet pole pair.
Driving Performance Estimation for Compact EV equipped with Cross-pole-Shape FR motors.
Abstract
This paper investigates the design and performance of cross-pole-shaped flux reversal (FR) motors intended for compact electric vehicle (EV) applications. The study first examines the influence of cross-pole width and identifies an optimal value of 2.0°, which enables high torque output while reducing magnet eddy current loss and suppressing even-order harmonics in the no-load induced voltage. An 18/24-pole FR motor is then optimally designed using a genetic algorithm. The optimized motor exploits the cross-pole-shaped stator to generate both magnet torque and reluctance torque, resulting in improved torque characteristics and a maximum torque of approximately 61 N·m. Performance comparisons with a previously reported in-wheel switched reluctance (SR) motor show that the FR motor delivers higher torque over the entire operating range and maintains efficiency above 90% across wide operating regions. Furthermore, vehicle-level evaluations assuming in-wheel motor application demonstrate superior driving performance, including excellent hill-climbing capability and a maximum vehicle speed of 55 km/h. This maximum speed is 10 km/h higher than that of a comparable compact EV equipped with SR motors, highlighting the advantages of the proposed FR motor for in-wheel EV propulsion systems.
Speed – Torque Range Expansion of In-Wheel Axial-Flux SR Motor for Compact EV
Abstract
In a previous paper, axial-flux-type switched reluctance motors (AFSRMs) were prototyped and installed in two rear wheels of a compact electric vehicle (EV). The prototype AFSRM was confirmed to drive the prototype EV. However, the drive range of the prototype EV was not sufficient since the instantaneous phase torque distribution control (IPTDC) proposed in the previous paper has a small torque in the high-speed region. This paper presents an improved control method for increasing the speed – torque range of AFSRM and proves its effectiveness through simulations and experiments.
GP-DS
Limited Visits to Urban Amenities by Parents Revealed Through Mobility Data
Abstract
Low economic mobility remains a significant urban challenge. According to the OECD, it takes four generations for Japan’s lowest-income families to reach the average income level. Early-life experiences are crucial for children’s future economic success, highlighting the importance of parents’ access to rich urban amenities and interaction opportunities during the early childhood years. This study proposes a novel method to identify parents of young children using human mobility data and analyzes the unique behavioral patterns of parents. The findings reveal that parents tend to avoid living in central urban areas, preferring suburban areas. They travel shorter distances and spend more time at home with limited exploration of new places. Specifically, regarding food outlets, they frequently visit nearby supermarkets but tend to avoid distant restaurants. These behavioral patterns were found to be consistent among parents, suggesting that time and financial constraints associated with parenting may limit their opportunities for diverse urban experiences and social interactions. This research provides a foundation for evaluating the urban experiences of parents and children and for identifying amenities that require intervention to enhance the future economic prospects of low-income families.
BibAI: An AI-Powered BibTeX Cleaning Agent
Abstract
BibTeX files are widely used in academic publishing to store and manage bibliographic references in a structured and consistent format. However, maintaining this consistency becomes increasingly difficult as BibTeX collections grow in size, leading to duplicated entries, incomplete metadata, and heterogeneous formatting that must be corrected manually. To address this challenge, we developed BibTeX Cleaner, an LLM-powered tool that automatically standardizes and enhances bibliographic data. The system uses deterministic rules to resolve common formatting issues and connects to bibliographic databases such as DBLP, Semantic Scholar, and arXiv to fill missing fields and retrieve up-to-date official metadata. LLMs complement these rule-based components by analyzing ambiguous or incomplete entries and suggesting consistent metadata representations. A JSON mapping file also tracks renamed and merged entries to ensure reliability and transparency. Finally, the existing features and format settings can be customized through a YAML configuration file to support different project needs.
Purify: Purify Your Prose, Clarify Your Science - A Script-based Paper Editing Tool for Transparent and Context-Aware LLM Use in Scientific Manuscripts
Abstract
Purrify is a script-based scientific manuscript editing tool designed to enable transparent and reference-grounded use of large language models (LLMs) for improving paper readability. Unlike existing commercial paper editing systems that overwrite original sentences with LLM-generated outputs, Purrify preserves author control by disclosing all model-suggested modifications as explicit annotations, ensuring full provenance and edit traceability. The tool incorporates Context-Aware Texting (CAT), which constrains LLM behavior using a user-supplied list of paper abstracts identified by trusted DOIs, reducing hallucination risk and enforcing scientific consistency. Through lightweight string processing, Purrify detects segments enclosed by //text// for LLM proofreading and ++text++ for contextual citation injection, automatically aligning statements with relevant references when supported by the provided abstracts. Currently deployed on Amazon servers using DeepSeek-V3 through AWS Bedrock, Purrify demonstrates a reliable, transparent, and ethically compliant framework for LLM-assisted scientific writing. Future work includes optimizing latency using bundled abstract summaries and expanding structured context control for high-stakes scientific editing.
Creating Email Agent using Large Language Model
Abstract
We present Email Agents, an LLM-powered assistant that reduces inbox overload by automatically classifying Gmail messages and drafting replies. Using the Gmail API for message retrieval and Vertex AI (Gemini 2.0 Flash) for inference, the system assigns one of four labels—Important, Read, Events, or Ads—and applies the corresponding Gmail label. For high-priority emails, it also generates replies to drafts to streamline response work. A scheduled job runs the pipeline periodically; the current prototype processes the latest 10 unread emails and marks them as read. In a demo, the system successfully identified ad emails and produced practical drafts for important messages, with an estimated cost of only a few to ~10 JPY per 100 emails. Future work includes extracting tasks/events into Google Calendar and optimizing prompts and execution cost.
