To promote open science, publishing and sharing research data is recommended. To enhance the accessibility of research data, metadata about research data should be assigned; however, manually assigning metadata is costly. On the other hand, scholarly papers describe information about research data, i.e., meta-information, which could potentially be utilized for metadata generation. This paper explores the feasibility of obtaining meta-information by utilizing scholarly papers. We implemented two methods to obtain meta-information from text surrounding URLs that cite research data, and evaluate their performance. The first method extracts meta-information from input texts, and the second method classifies input texts. Both methods are performed using a LLM. The experimental results indicate that meta-information extraction using an LLM has low performance. By contrast, in research data classification, we confirm that the performance of our method improves by providing classification examples.

In recent years, image generation techniques using diffusion models have rapidly advanced. Their flexibility, realism, and ease of implementation have raised ethical concerns, such as the spread of fake news. Conventional approaches for detecting these generated images typically rely on applying computationally expensive neural networks to the entire high-resolution image. However, our previous research revealed that partitioning the image and focusing on local regions can achieve both higher detection accuracy and lower computational load. In this study, we investigate why such detection methods are effective and conduct a detailed examination of the characteristics of images generated by diffusion models.

Environmental Sound Classification (ESC) is a crucial technology for understanding surrounding conditions, and recently, methods utilizing the Vision Transformer framework have attracted attention. However, Transformer models are prone to overfitting when data is insufficient, and pre-trained models may not adapt well to the target sound environment. On the other hand, CNNs exhibit stable performance even without pre-training and with small amounts of data, offering the advantage of reducing the impact of noise through convolutional processing's denoising capabilities. Therefore, this study focuses on the noise resistance of CNNs and examines the selection of the optimal CNN and the introduction of data augmentation techniques. First, five proven CNNs were compared, and then the data augmentation technique CutMix was introduced to improve performance with noisy data. The results showed that EfficientNet exhibited excellent noise resistance, and CutMix improved overall classification performance. These findings contribute to the practical application of high-accuracy and noise-resistant ESC models.

Stereophonic sound is widely used in movies, games, etc., and can provide viewers with a more realistic and immersive experience. We constructed a system to generate stereophonic sound from performance videos by combining object recognition and sound source separation.The generated stereophonic sound was evaluated qualitatively by questionnaires and quantitatively by comparing acoustic features.The results show that the system can generate stereophonic sound without the need for special equipment or expert engineers.It was also found that the system can be further improved by revising the positioning of the sound sources.

Choosing an appropriate font is crucial in graphic design for advertisements and posters. However, conventional font search methods require users to sift through many fonts to find one that meets their needs, demanding considerable effort before arriving at the right choice. To address this issue, we propose an interactive font search system that leverages a large-scale language model and utilizes textual and visual information. In our proposed method, users retain the freedom to browse fonts sequentially while engaging in intuitive chat-based interactions that effectively capture their intentions, enabling a more efficient font search. Through user studies, we compared our system with a baseline method and demonstrated that our approach can more effectively identify fonts suited to the target advertising images.

Brain-computer interface (BCI) is technology to control computers directly using the signal from brain and is expected to have various applications in rehabilitation and daily life support, especially for people with severe motor dysfunction. Low computational cost and high generalization performance are important for the widespread use of BCI. In this study, we propose the CNN-GRU hybrid model to reduce computational cost and improve generalization performance using domain adaptation. We investigated the discrimination performance of opening/closing either the left or right hand using the dataset OpenBMI as target domain data and the other four datasets (Physionet, Kaya2018, Meng2019, and Stieger2021) as source domain data. The proposed frameworks achieve an accuracy of 72.2%. At the same time, 32 out of 54 participants reached the accuracy of 70%, BCI control, which assumes binary classification.

As the development of services utilizing large-scale language models (LLMs) continues, there is growing interest in the use of AI technology in the mental health field as well. We have been developing a chatbot using LLM as part of a service to support individual self-care. However, there are issues such as the lack of consistent conversation and personalized responses. In this study, we summarized the issues of the current chatbot based on the opinions received from users and developed a memory system to realize “responses utilizing past conversations” among them. The memory system implemented in this study enabled the generation of suggestions, topics, and answers to user questions based on the content of past conversations. This paper describes the design and implementation of the memory function using a large-scale language model and shows an example response.

Accurately and continuously recognizing users' emotional states in real-time is essential for enabling a dialogue system to adapt to users flexibly. This becomes particularly challenging when speech and linguistic cues are unavailable, such as when the user is not taking a turn. In such cases, non-verbal information becomes crucial in understanding user emotions. In this study, we aimed to develop a model that classifies users' emotional valence during conversations in real-time using physiological signals. Specifically, we utilized multimodal dialogue data, including physiological signals, collected in our previous research. We attempted to build a model that estimates the user's emotional valence, categorized as positive or negative, by leveraging a time-series model on arbitrary segments of physiological signals (EDA, BVP, and PPG) recorded during the dialogue. Experimental results demonstrated that integrating multiple physiological signals enhances emotion estimation performance, highlighting the potential of physiological data to improve real-time emotion recognition in dialogue systems.

“Nudge theory”, known in the field of behavioral economics, is applied to methods of encouraging users to use objects appropriately. However, nudges are scarcely versatile because of the difference depending on the type of objects to implement. On the other hand, It has been suggested that “anthropomorphization” changes user’s attitude toward objects. Anthrpomorphizing which includes verbal instructions is implemented a lot in prior researches, but it can be versatile methodology when we implement it only visual elements. The purpose of this study is to develop “anthropomorphizing device” and to verify the validity of it. In the subject experiment, we used pen holders as the object and compared it with a device implemented according to nudge theory and a device displaying unrelated plane figures to confirm whether or not the anthropomorphic device “the effect of making it easier to put a pen in the pen holder”. The results showed that the pen holders with anthropomorphic devices were more likely to attract attention than other pen holders, but did not show a significant difference in the effect of the pen holders on pen insertion.

The quality inspection of requirement specifications for software products is quite important for reducing development costs and maintaining product quality in the development process. Currently, this task is performed manually, which requires significant effort and has risks to make oversights and errors. To address these issues, we utilize large language models in conjunction with a predefined metamodel of software requirements specification to structure the content of a given requirement specification into a knowledge graph and assess its semantic quality. We report on the results of evaluating the effectiveness of the proposed approach by applying it to actual requirement specifications and comparing them with the results of manual quality inspection.

In portfolio management, directly using past stock returns as training data often results in suboptimal performance due to covariate shifts. To address this, deep reinforcement learning (DRL) has emerged as a powerful approach that dynamically adapts to market conditions and optimizes long-term rewards while reducing reliance on short-term price fluctuations. However, the black-box nature of DRL poses a significant challenge to transparency, which is crucial for financial institutions to meet regulatory and accountability requirements. Counterfactual explanations provide a promising solution for enhancing the interpretability of DRL-based investment strategies. However, a key challenge is ensuring that counterfactual scenarios maintain consistency between stock prices and their derived technical indicators. To address this, we propose a novel two-step framework: first, we identify critical stocks that significantly influence the agent’s decisions, and then we systematically adjust their prices to generate realistic counterfactual scenarios. Experimental evaluations demonstrate that the proposed method improves transparency in DRL-driven portfolio management, offering a more interpretable and robust framework for investment decision-making.

In the design process of press forming dies, the finite element method (FEM) is widely used to obtain die shapes that satisfy formability and dimensional accuracy of pressed parts. However, high computation costs and increasingly complex part geometries have prolonged the cycle of modifying die designs based on simulation results and re-analysis, hindering overall efficiency. To address these challenges, we propose a method to construct highly accurate surrogate models based on FEM outcomes. Specifically, we apply PointNeXt, a deep learning model originally developed for 3D point cloud shape classification and semantic segmentation, to learn the relationship between press die and the resulting pressed part shapes. The surrogate model can predict in minutes with an average error of approximately 0.3 mm, meeting practical accuracy requirements. This approach is expected to streamline the design process, shorten development cycles, and reduce lead times in die design.

We develop price sentiment indices (PSI) to analyze price trends from both consumer and corporate perspectives. We extract price-related comments from the Economy Watchers Survey and classify their price direction using pre-trained language models including large language models (LLMs). We show that combining FinBERT and LLM can classify the direction of prices more efficiently and more accurately. We then develop detailed PSIs from the perspectives of consumer and business, as well as goods and services, by categorizing comments based on the domain and the respondent's industry. The PSIs show higher correlations with existing indices compared to previous studies. Notably, the PSIs for consumers exhibit even stronger correlations.

In this study, we discuss the applicability of Quantum Circuit Born Machine (QCBM) as surrogate models for simulations involving complex stochastic variations. It is well-known that QCBM can learn complex probability distributions, such as Gaussian mixture distributions, through the probability amplitudes of quantum circuits. However, research on applying trained QCBM as sampling method for probabilistic physical simulations has not been sufficiently investigated. Therefore, this paper focuses on the sampling accuracy and speed of QCBM and evaluates their feasibility through a comparison with rejection sampling.

Online education has rapidly advanced, offering recognized convenience and flexibility. However, effective methods to monitor student concentration remain a challenge. This study develops a feedback system combining gaze tracking and head movement recognition to enhance online class quality and teacher-student interaction. A 30-minute simulated online class with five participants acting as "focused" and "distracted" students was conducted via Zoom. Results showed "focused" students had lower gaze-away rates, while "distracted" students had higher rates, aligning with their roles. The system demonstrated its ability to accurately evaluate concentration and its potential to improve online education. Future work will focus on broader validation and system refinement.

The gas chromatography-mass spectrometer (GC-MS) is an instrument capable of measuring the concentrations of various volatile substances. It is used as input data for multivariate analysis and machine learning applications across a wide range of fields from quality control of materials and food to disease diagnosis. To utilize the results of this multivariate analysis and machine learning as scientific insights, it is essential to identify the substance names from the contributing dimensions. Typically, the identification process involves searching a library of standard spectra using the similarity in mass distributions (spectra) of molecular fragments obtained by applying high energy to the molecules. However, due to the presence of many similar spectra, numerous candidate substances exist, and the correct compound may be overlooked. This paper proposes a system to improve the workflow by using knowledge processing technology. Practical examples using open data related to the quality of materials and foods demonstrate a reduction in missed compounds.

In industrial inspections such as visual and sensory inspections in the manufacturing industry, the automation of industrial inspection using computer vision technology is expected to reduce costs and minimize variations caused by human factors. For industrial products, where anomalies are rare and unknown anomalies may occur, unsupervised anomaly detection methods such as PaDiM and PatchCore have been proposed to detect anomalies based on normal products only. However, the models used in these methods are pre-trained on datasets like ImageNet and thus do not learn the features of the inspection target. This makes it difficult to adjust the models for performance improvement. Therefore, this paper proposes an anomaly detection method that uses self-supervised learning to capture the features of images targeted. Specifically, the model is fine-tuned with inspection images using self-supervised learning, and anomaly detection is performed using an unsupervised anomaly detection method with the fine-tuned model. In experiments, the proposed method achieved higher image-level AUC than existing methods for the visual inspection of breaker nameplates.

In recent years, numerous large-scale language models have been proposed and released. Meanwhile, smaller-scale language models, which can run in environments with comparatively limited resources, have also been made available and are being utilized. Demand for these models is rising, particularly for security reasons such as the need to avoid sending data outside the system, as well as in scenarios where network connectivity is unstable—like onboard aircraft or during high-speed travel.In this study, we introduce the quantization and fine-tuning techniques required to effectively utilize small-scale language models (SLMs) under such constrained conditions, and present our evaluation of how these approaches impact model accuracy.

A module for proposing plant operation using reinforcement learning has been developed that quickly proposes appropriate operations depending on the state of the plant. However, there was a tendency to propose operations that were difficult to implement in actual operation, such as proposing upper and lower limits for the range of possible operations. In this study, we applied RLHF that incorporates human feedback as a reward and developed a module that can propose more realistic and safer operations.

We investigated the use of deep learning-based object detection models for identifying estrus behavior in breeding cattle from video. In livestock reproduction, accurately determining the timing of artificial insemination is critical for efficient operations. Therefore, it is highly desirable to develop automated techniques to detect estrus-related behaviors, such as mounting, in surveillance camera footage. This study explores the detection of mounting behavior using existing object detection algorithms. Specifically, we compared the performance of a lightweight model (YOLO, based on CNN) optimized for computational efficiency and a high-accuracy model (DETR, employing a Transformer architecture). Our objective was to provide insights into optimal model selection for video-based monitoring of breeding cattle. Experimental comparisons using farm data demonstrated that the DETR model exhibited superior detection performance for mounting behavior, even under conditions with partial occlusion.

Mechanistic Interpretability (MI) aims at the causal interpretation of the Language Models (LMs). Typically, MI studies explore a circuit representing a specific concept by deactivating its components or mimic LM’s inner workings by replacing the activation with a white-box algorithm. Although these approaches provide valuable insights for LM’s inner workings (intrinsic causality), the inner workings are affected by the causality of the problem they are addressing, or extrinsic causality. To understand this dual causality, I provide the formal framework for estimating the causal effect of extrinsic causality over intrinsic causality. Then I reframe my previous findings into this dual causality problem; the bias of pretraining on the image-text interaction in three Transformer variants performing hateful memes detection. The result shows that our framework can quantify the pertaining bias affects the entire causal relationship of this problem and a hypothetical causal effect where only a subset of the attention matrix causes the model output. In conclusion, this paper contributes to MI communities by providing the initial findings about a missing component in MI; interaction with the world.

This paper describes an analysis of physical activities in older adults engaged in rhythm game-based eSports using a multimodal data approach. While eSports are increasingly recognized as a viable recreational activity for older adults, methods for comprehensively assessing their physiological and behavioral responses remain insufficiently developed. To address this gap, we constructed a data collection framework that integrates heart rate monitoring, in-game event tracking, and gaze and head movement analysis. The results suggest that playing eSports induces measurable physiological and behavioral changes, highlighting the potential of multimodal data analysis to better understand older adults' engagement in eSports.

This paper describes an attempt to extract clinical questions from a clinical registry database using Bayesian Networks. We generated Bayesian network models using clinical registry data from J-DREAMS, and showed that it is possible to generate clinical questions from the node pairs by extracting combinations of nodes with large amounts of mutual information for a specific node through sensitivity analysis.

Artificial Neural Networks (ANN) involve significant computational costs during updates due to catastrophic forgetting, where new learning overwrites existing knowledge, necessitating the relearning of all data. Despite various attempts to address this issue, no method has yet reached a practical level. Brain-inspired methods are particularly noteworthy, as biological neural networks do not suffer from catastrophic forgetting. Researchers have attempted to mimic the brain's memory mechanisms, achieving state-of-the-art performance in some cases. However, the developmental aspects of the memory mechanisms have not been sufficiently considered. By examining the brain's functional developmental sequence and the stimuli necessary for development, we aim to gain insights into suppressing catastrophic forgetting. This paper introduces the experimental setup and shares preliminary results.

The aim of this study was to classify initiatives presented at the Digi-den Koshien (National Championship for Achieving a Digital Garden City Nation) using Latent Dirichlet Allocation and to provide an overview of their trends. The Digi-den Koshien is held to promote the Vision for a Digital Garden City Nation by soliciting examples of local problem solving using digital technologies from companies and organizations, and awarding outstanding initiatives to encourage their further development. Based on the textual content of the initiatives, a topic model was constructed. The characteristics of the award-winning initiatives were then examined using the model.