Published papers

We present prospects for the $Θ^+$ pentaquark baryon search using the newly constructed LEPS2 facility at SPring-8. The LEPS2 detector system features significant improvements in acceptance for multi-particle final states compared to previous experiments. Our search employs two complementary strategies: direct production in the $γn \to K^-Θ^+$ reaction using a liquid deuterium target with a photon beam up to 2.4 GeV, and $\bar{K}^{*0}$-associated $Θ^+$ production using a liquid hydrogen target with a photon beam up to 2.9 GeV. The extended acceptance covers both forward and large angle regions, effectively spanning the kinematic regions explored by previous LEPS and CLAS experiments. The large acceptance and improved resolution of LEPS2, combined with these complementary approaches, provide unprecedented sensitivity for establishing the existence of the $Θ^+$ or placing definitive upper limits on its production.

Our previous work established the feasibility of FPGA-based real-time trajectory anonymization, a critical task for protecting user privacy in modern location-based services (LBS). However, that pioneering approach relied exclusively on shortest-path computations, which can fail to capture re- alistic travel behavior and thus reduce the utility of the anonymized data. To address this limitation, this paper introduces a novel, history-aware trajectory k-anonymization methodology and presents an advanced FPGA-based hardware architecture to implement it. Our proposed architecture uniquely integrates par- allel history-based trajectory searches with conventional shortest- path finding, using a custom fixed-point counting module to ac- curately weigh contributions from historical data. This approach enables the system to prioritize behaviorally common routes over geometrically shorter but less-traveled paths. The FPGA implementation demonstrates that our new architecture achieves a real-time throughput of over 6,000 records/s, improves data retention by up to 1.2% compared to our previous shortest-path- only design, and preserves major arterial roads more effectively. These results signify a key advancement, enabling high-fidelity, history-aware anonymization that preserves both privacy and behavioral accuracy under the strict latency constraints of LBS.

Digital twins (DTs) are virtual replicas of physical assets, systems, and processes that enable real-time monitoring, simulation, and analysis to improve decision-making, optimize operations, increase efficiency, and achieve cost savings across various Internet of Things (IoT) applications. This paper introduces an IEEE 1451-based smart transducer digital twin framework (DTF) for IoT real-time monitoring and control applications. The framework consists of an IEEE 1451-based smart transducer physical twin (PT), its corresponding digital twin (DT), and a bidirectional information exchange between them for real-time monitoring and control. In this paper, the IEEE 1451.0 and P1451.1.6-based smart transducer DT is developed using the Node-RED platform and message queue telemetry transport (MQTT) protocol. The PT is developed using embedded devices, and a sensor and actuator development kit. The PT and DT communicate using 1451.0 and P1451.1.6 messages to exchange temperature sensor data for controlling the temperature of a smart home in real-time. The monitoring and control results of the smart home are provided in the paper to test and validate that the physical twin and digital twin work correctly.

Recently, smart agriculture utilizing Information and Communication Technology (ICT) and robotics has gained attention. The Institute of Electrical and Electronics Engineers (IEEE) Standards Association is developing standard P2992 for smart agricultural data design. Research on crop feature representation exists but lacks universal methods and real-world application. This study proposes a method to represent crop growth and environments in JavaScript Object Notation (JSON) format based on P2992, applied to cooperative farmers’ greenhouse management. The method includes feature extraction, environment data acquisition, and JSON representation. As a result of the designed application, flower detection achieved 84.5% conformance and 86.1% recall, with a data size reduction of over 99% compared to traditional methods. Applications using temperature and crop data show potential for effective cultivation and sales management by using simple JSONata scripts, suggesting P2992's practicality and usefulness in real-world environments.

Multi-task learning has become increasingly popular in the machine learning field, but its practicality is hindered by the need for large, labeled datasets. Most multi-task learning methods depend on fully labeled datasets wherein each input example is accompanied by ground-truth labels for all target tasks. Unfortunately, curating such datasets can be prohibitively expensive and impractical, especially for dense prediction tasks which require per-pixel labels for each image. With this in mind, we propose Joint-Task Regularization (JTR), an intuitive technique which leverages cross-task relations to simultaneously regularize all tasks in a single joint-task latent space to improve learning when data is not fully labeled for all tasks. JTR stands out from existing approaches in that it regularizes all tasks jointly rather than separately in pairs -- therefore, it achieves linear complexity relative to the number of tasks while previous methods scale quadratically. To demonstrate the validity of our approach, we extensively benchmark our method across a wide variety of partially labeled scenarios based on NYU-v2, Cityscapes, and Taskonomy.

In IoT applications, service provision commonly leverages the cloud infrastructure. However, the processing system at the edge is expected recently. It can process private information collected in factories and other locations without exposing it externally by using a distributed networked system. Typically, this private information is encrypted, and it can only be decrypted in the cloud. It has been difficult to provide the processing system at the edge. This study proposed and implemented the system to enable secure, Transport Layer Security (TLS) 1.3-compliant, and authentication-compliant processing under an encrypted communication environment between the terminal and the cloud, based on the system to realize edge processing called a transparent add-on. It also defines communication protocols. The latency introduced by the introduction of the proposed system is equivalent to the delay caused by adding one router to the network path. Therefore, the proposed system does not cause any harm to the provision of services after it is implemented.

With the enhancement of personal information protection regulations in many countries, there is a growing need to develop an opt-in/opt-out system that allows users to specify the recipients and anonymization level of their information, as well as the information distribution channels. A previous study proposed a dataflow management system to perform fast route selection processing on the communication path by adding tags to packets. However, prior research assumes fixed IP network and does not support mobile networks, in which IP addresses are dynamically assigned. In this study, we examine the requirements for a dataflow management system on mobile networks, propose a cache on a network node (in-network cache), and a cache migration method corresponding to user mobility. We conducted simulations using ns-3 under both simplified and real-world mobility scenarios. The results demonstrate the effectiveness of our proposed method in reducing communication overhead and enhancing network robustness.

Recently, the increased utilization of data has led to the development of various services. However, using data necessitates anonymization techniques for privacy. multi-input k-anonymizer unit (MIKU)[1], a facial image anonymization method proposed by Nakamura et al., anonymizes images by averaging the latent vectors of facial images in the latent space of StyleGAN[2]. However, MIKU has two limitations: (1) it only targets single-face images, and (2) it merely calculates the centroid of the latent vectors between the original images, thus not preserving the statistical information of the original image dataset. Therefore, this paper proposes a framework to realize a k-anonymization method for photos containing multiple faces. The anonymization process uses a proposed framework comprising multiple modified artificial intelligence applications. First, identifiers, quasi-identifiers, and sensitive attributes are specified for a photo with multiple faces. This paper proposes three methods for generalizing age, one of the quasi-identifiers. The first is the random method, which replaces the age of the original image with the age of a random individual. The second is the uniform distribution method, which assigns ages based on the uniform distribution of age ranges after anonymization. The third is a kernel method that assigns ages using random numbers obtained from a probability density function computed by kernel density estimation. The experiments resulted in an Fréchet inception distance (FID)[2] equivalent to MIKU without any observed difference for the specified sensitive attributes. Furthermore, the most effective results were obtained using the kernel method, indicating that the proposed method is effective for the k-anonymization of facial images.

This paper describes a collision avoidance method for a biped robot with an upper body. We propose a method wherein the robot stops in front of an obstacle by generating arm force. When the robot detects the obstacle by the arm tip, it should stop short of the obstacle to avoid crash. Hence, we propose trajectory planning in consideration of the pushing force of the arm. The arm force is controlled to be generated as a function of the distance from the robot body to the obstacle. The closer the robot approaches the obstacle, the larger the arm force becomes. As a result, the robot can stop by utilizing the arm force. In case the obstacle is unmovable, the robot can stop by exerting arm force. If it is movable, the robot can continue walking by pushing it. In this paper, the linear inverted pendulum mode (LIPM) and the idea of orbital energy are introduced, and then, we extend LIPM and orbital energy in consideration of the dynamics of the arm force. The extended orbital energy is utilized to discriminate whether the robot can stop or not and to modify the trajectory of the robot to avoid collision. © 2007 IEEE.

In recent years, there has been growing interest in using big data for machine learning applications. However, as data are often distributed across multiple clients, privacy and physical limitations can make it difficult to aggregate and use these data on centralized servers. Federated learning and split neural networks enable machine learning models to be built without requiring the aggregation of distributed data. These methods have received increased attention in the industrial sector as a means of eliminating impediments to privacy preservation in data sharing across various institutions, such as smart factories. One critical aspect of implementing such methods in practice is analyzing the time variation of traffic and estimating the throughput required for operationalizing machine learning. Previous studies have formulated traffic characteristics without considering packet overhead and measured only the total traffic and the throughput per iteration; such an approach is inadequate from a traffic monitoring perspective. To address these issues, this study conducted an empirical investigation of traffic patterns under realistic conditions, with a particular focus on quantifying traffic and latency. The results of this study suggest that due to the dynamic nature of throughput, it is essential to measure traffic using time units that can accommodate the temporal variability of traffic.

We propose EAGLE update rule, a novel optimization method that accelerates loss convergence during the early stages of training by leveraging both current and previous step parameter and gradient values. The update algorithm estimates optimal parameters by computing the changes in parameters and gradients between consecutive training steps and leveraging the local curvature of the loss landscape derived from these changes. However, this update rule has potential instability, and to address that, we introduce an adaptive switching mechanism that dynamically selects between Adam and EAGLE update rules to enhance training stability. Experiments on standard benchmark datasets demonstrate that EAGLE optimizer, which combines this novel update rule with the switching mechanism achieves rapid training loss convergence with fewer epochs, compared to conventional optimization methods.

As the quantities of data recorded by embedded edge sensors grow, so too does the need for intelligent local processing. Such data often comes in the form of time-series signals, based on which real-time predictions can be made locally using an AI model. However, a hardware-software approach capable of making low-latency predictions with low power consumption is required. In this paper, we present a hardware implementation of an event-graph neural network for time-series classification. We leverage an artificial cochlea model to convert the input time-series signals into a sparse event-data format that allows the event-graph to drastically reduce the number of calculations relative to other AI methods. We implemented the design on a SoC FPGA and applied it to the real-time processing of the Spiking Heidelberg Digits (SHD) dataset to benchmark our approach against competitive solutions. Our method achieves a floating-point accuracy of 92.7% on the SHD dataset for the base model, which is only 2.4% and 2% less than the state-of-the-art models with over 10× and 67× fewer model parameters, respectively. It also outperforms FPGA-based spiking neural network implementations by 19.3% and 4.5%, achieving 92.3% accuracy for the quantised model while using fewer computational resources and reducing latency.

As the volume of data recorded by embedded edge sensors increases, particularly from neuromorphic devices producing discrete event streams, there is a growing need for hardware-aware neural architectures that enable efficient, low-latency, and energy-conscious local processing. We present an FPGA implementation of event-graph neural networks for audio processing. We utilise an artificial cochlea that converts time-series signals into sparse event data, reducing memory and computation costs. Our architecture was implemented on a SoC FPGA and evaluated on two open-source datasets. For classification task, our baseline floating-point model achieves 92.7% accuracy on SHD dataset - only 2.4% below the state of the art - while requiring over 10x and 67x fewer parameters. On SSC, our models achieve 66.9-71.0% accuracy. Compared to FPGA-based spiking neural networks, our quantised model reaches 92.3% accuracy, outperforming them by up to 19.3% while reducing resource usage and latency. For SSC, we report the first hardware-accelerated evaluation. We further demonstrate the first end-to-end FPGA implementation of event-audio keyword spotting, combining graph convolutional layers with recurrent sequence modelling. The system achieves up to 95% word-end detection accuracy, with only 10.53 microsecond latency and 1.18 W power consumption, establishing a strong benchmark for energy-efficient event-driven KWS.

The Internet of Things (IoT) is highly heterogeneous regarding smart sensors/devices, connectivity, communication protocols, and data formats. The major challenges of IoT ecosystems are fragmentation (or disintegration) and cross-domain interoperability. To overcome the challenges of interoperability, standardized interfaces and protocols, harmonized interactions, and interoperability testing are essential. This paper introduces a harmonization model of interactions among IoT devices with diverse interfaces/connectivity and communication protocols to achieve interoperability by harmonizing their interactions. A harmonization implementation of interactions of the IEEE 1451-based IoT devices and MODBUS devices has been provided in the paper to test and verify that the proposed harmonization model works well. This harmonization model and implementation of interactions among IoT devices will provide a solid foundation for the IEEE P1451.99 standard specification.

Short-term hot water consumption forecasting in an individual household is important to realize energy-efficient hot water management while meeting users' comfort. However, consumption traits in an individual household contain much irregularity, and that gives additional challenges to accurate forecasting. This paper proposes an Attention-PVS model to tackle this issue, which focuses on consumption values with similar past consumption trends. In this study, verification experiments were conducted on a real-household hot water consumption dataset collected by Electricity of France. The model was evaluated with its consumption forecasting error with MAE, MSE, and RMSE metrics and computational costs. As for the computational cost, experiments were conducted on NVIDIA Jetson Nano (Jetson) to validate the applicability to embedded systems. The results revealed that the proposed model performs consumption forecasting with comparable accuracy to other state-of-the-art models. Additionally, while LSTM scored lower error than the proposed model, Attention-PVS performed training and inference on Jetson in shorter times than LSTM.

To prevent personal information leakage from the collection and publication of big data, anonymization techniques that protect privacy while preserving data utility are essential. In particular, m-invariance provides privacy protection for dynamic datasets that are continuously published as records are updated. However, previous studies have faced challenges where execution time increases with the total dataset size, and no evaluations have been conducted for large-scale datasets. In our replication of prior research, the anonymization of 1 million dynamic records required an average execution time of approximately 7 h. To address this issue, we propose a novel m-invariance algorithm focused on record insertion and deletion, which is independent of the total dataset size. We implemented our algorithm on both CPU and FPGA platforms, achieving execution times of under 5 s for 5000 updates in a 1 million-record dataset on both platforms. The throughput per updated record was measured at 57 ms/record for the CPU-based implementation and 2 ms/record for the FPGA-based implementation. Furthermore, through buffer resource optimization on FPGA, we demonstrated the capability to anonymize large-scale dynamic datasets exceeding 1 million records. In summary, this study achieves m-invariance anonymization for 1 million dynamic records, enabling updates at a rate of 1 s/record. © 2025 The Author(s). IEEJ Transactions on Electrical and Electronic Engineering published by Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.