| YIN FEIYANG | M, 1回目発表 | インタラクティブメディア設計学 | 加藤 博一 | 向川 康博 | 澤邊 太志 | Isidro Butaslac | |
|
title: Making the Invisible Visible: Exploring How Micro-Expressions Impact Empathy in Face-to-Face communication
abstract: When people engage in face-to-face communication, visible facial expressions serve as one of the most immediate and informative parts of emotional cues, helping us interpret emotional states, foster emotional alignment and mutual understanding, and ultimately enhance empathy. However, what can typically be perceived in such interactions are only Macro-Expressions. Due to the short duration and low intensity, Micro-Expressions are usually imperceptible or neglected by the naked eyes. As a result, they remain undetected in natural interactions. Hence, this research attempts to visualize Micro-Expressions, in order to investigate their influence on empathy. language of the presentation: English | |||||||
| FAKIH ABDEL HAMID | M, 1回目発表 | サイバーレジリエンス構成学 | 門林 雄基 | 安本 慶一 | 妙中 雄三 | ||
|
title: Pridwen: A Privacy-Preserving Offline AI Framework for Gamified Cybersecurity Education
abstract:Digital illiteracy and insufficient cybersecurity training continue to expose non-technical users such as administrative and HR staff to phishing, unsafe interfaces, and manipulation. These risks remain critical, as phishing and social engineering persist as dominant attack vectors, and nearly two billion people still lack reliable internet access. Existing tools are typically reactive, cloud-dependent, overly complex for beginners, and unable to adapt to diverse cultural contexts, limiting privacy, accessibility, and sustained engagement. This study presents Pridwen, an initial effort toward developing a lightweight, AI-driven, gamified, and fully offline platform for proactive cybersecurity education. The system integrates a local LLM, scenario-based simulations, and an adaptive five-module curriculum. Through this platform, we aim to examine whether combining adaptive instruction with gamification and fine-tuned local LLM guidance can enhance learning outcomes, usability, and engagement across varied demographic and cultural groups, with a focus on non-technical users. A pilot evaluation employing pre/post quizzes, SUS scores, engagement metrics, and statistical testing will clarify both the limitations and the educational potential of Pridwen. language of the presentation:English | |||||||
| 谷口 蒼馬 | M, 1回目発表 | サイバーレジリエンス構成学 | 門林 雄基 | 林 優一 | 妙中 雄三 | ||
|
title: Proposal of a CAPTCHA Authentication Method Applying the Characteristics of Large Language Model Abstract: In recent years, with the rapid advancement of Large Language Models (LLM), conventional image-recognition CAPTCHAs have become easily breached by bots, diminishing their effectiveness. In this study, we propose a new CAPTCHA authentication method that distinguishes between humans and bots by applying the safety guardrails inherent in LLM. Specifically, we embed harmful prompts, to which LLM refuses to respond, into the CAPTCHA authentication process. This approach is based on the hypothesis that while humans follow the normal authentication process, bots read the prompts, trigger their guardrails, and consequently fail the authentication. Unlike traditional CAPTCHA approaches that exploit the performance limitations of LLM, we aim to establish a new CAPTCHA authentication technology by applying the inherent characteristics of LLMs for defense. language of the presentation: Japanese 発表題目: 大規模言語モデルの特性を応用したCAPTCHA認証の提案 発表概要: 近年、大規模言語モデル(LLM)の急速な発展に伴い、従来の画像認識CAPTCHAはボットによって容易に突破されるようになり、その有効性は低下している。本研究では、LLMが持つ安全性に関するガードレールを応用し、人間とボットを判別する新しいCAPTCHA認証手法を提案する。具体的には、CAPTCHA認証にLLMが応答を拒否する有害なプロンプトを埋め込む。人間は通常の認証プロセスを辿る一方で、ボットはプロンプトを読み取ることでガードレールが作動し、認証に失敗するという仮説に基づく。従来のCAPTCHA認証のように、LLMの性能の限界を利用するアプローチとは異なり、その特性そのものを防御に応用することで、新たなCAPTCHA認証技術の確立を目指す。 | |||||||
| 谷 昌俊 | M, 1回目発表 | サイバネティクス・リアリティ工学 | 清川 清 | 和田 隆広 | 内山 英昭 | Perusquia Hernandez Monica | 平尾 悠太朗 |
|
title: Investigation of the Effects of Active Control on Vection Intensity in Underwater VR
abstract: Vection is a visually induced illusion of self-motion and has been widely studied as a key factor in enhancing the sense of presence and immersion in virtual environments. Recent studies have reported that underwater environments generate different self-motion perception compared to terrestrial environments due to reduced gravitational sensation caused by buoyancy. This phenomenon may contribute to the realization of pseudo-infinite swimming in underwater VR. However, most previous studies on underwater vection have focused solely on passive visual stimulation, and the effects of combining such stimuli with active control have not been sufficiently examined. Therefore, this study aims to comparatively investigate the effects of active control methods, such as joystick operation and in-place swimming motions, on vection intensity in underwater environments. language of the presentation: Japanese 発表題目: 水中VRにおける能動的操作がベクション強度に及ぼす影響の検討 発表概要: ベクションは自己運動感覚を誘発する視覚的錯覚であり，仮想環境における臨場感や没入感を高める要素として広く研究されてきた．近年，水中環境では浮力により重力感覚が軽減されるため，地上とは異なる自己運動知覚が生じることが報告されている．この現象は水中VRにおける擬似的な無限遊泳の実現に寄与する可能性がある．しかし，従来の水中ベクション研究の多くは受動的な視覚刺激のみを対象としており，能動的な操作との組み合わせによる影響は十分に検討されていない．そこで本研究では，ジョイスティック操作やその場泳ぎといった能動的操作が水中環境におけるベクション強度に与える影響を比較検討する． | |||||||
| 中野 篤史 | M, 1回目発表 | ソーシャル・コンピューティング | 荒牧 英治 | 安本 慶一 | 若宮 翔子 | 久田 祥平 | |
|
title: Estimating Users’ Trust in AI Based on Questions About Their Medical Conditions abstract: In the medical domain, where the use of diagnostic AI continues to advance, users’ trust in such AI systems is a crucial factor. Traditional trust research primarily relies on questionnaire methods, but these impose a substantial burden in practical settings because users must respond manually. This study therefore explores a method for estimating user trust using the questions they input to the AI system. We prepared scenarios for three medical conditions (bone fracture, diabetes, and mental disorders) and collected users’ question texts and trust scores through crowdsourcing. These data were analyzed using a regression model based on Japanese BERT. The results indicated that, although there are limitations to fully estimating trust levels, it is possible to capture certain tendencies from textual input. Future work should focus on constructing QA systems that take user trust levels into account. 発表題目: 疾患に関する質問から AI への信頼感を推定する 発表概要: 診断 AI などの利活用が進む医療分野において、医療 AI に対するユーザーの信頼感は重要 な要因である。従来の信頼感研究では質問紙法が主に用いられているが、利用者がアンケー トに回答する必要があるため実運用では負担が大きい。そこで本研究では、ユーザーが入力 した AI に対する質問文を用いて信頼感を推定する手法を検討した。材料として、三疾患(骨 折・糖尿病・精神疾患)を対象にシナリオを設定し、クラウドソーシングでユーザーの質問 文と信頼感スコアを収集した。これらを日本語 BERT による回帰モデルで分析した。結果 は、信頼感の推定には限界があるものの、テキストから一定程度の傾向を捉えることが可能 であった。今後は、ユーザーの信頼度を考慮した QA システムの構築が求められる。 | |||||||
| XU JINSHA | M, 1回目発表 | ソーシャル・コンピューティング | 荒牧 英治 | Sakriani Sakti | 若宮 翔子 | PENG SHAOWEN | |
|
title: Probing Cultural Differences in LLMs Based on Loneliness Understanding abstract: Large language models (LLMs) such as Qwen are increasingly used in cross-cultural contexts, yet their cultural sensitivity remains unclear. This study examines cultural differences in Qwen-based loneliness assessment by evaluating a Japanese dialogue dataset across three prompt languages (Japanese, English, Chinese) with varied persona settings. Results show that Qwen fails to reliably distinguish fine-grained loneliness types—especially social recognition (SR)—exhibits large, conservative scoring errors with regression toward the mean, and is strongly affected by prompt language, whereas nationality and gender personas have only minor effects. These findings reveal that, at least in the case of Qwen, the model does not achieve a deep or stable understanding of Japanese loneliness and underscore the need for culturally grounded alignment in LLMs. language of the presentation: English | |||||||
| 白川 綾音 | M, 1回目発表 | ソフトウェア設計学 | 飯田 元 | 井上 美智子 | 柏 祐太郎 | Reid Brittany | |
|
title: Empirical analysis of consecutive test failures (Test alert snooze) in continuous integration
abstract: Continuous Integration (CI) automatically executes tests with every software change. This allows the software to be constantly maintained in a releasable state, thereby shortening the interval before it can be delivered to users (the release cycle). While tests are ideally always successful, in real-world development, they often fail multiple times. If a new defect is introduced while CI tests are already failing, this new defect can be hidden among the existing test failures. Consequently, the discovery of the newly introduced defect is delayed, and the time required for its correction increases. In this research, to make the defect fixing process more efficient, we define a test that has failed two or more consecutive times as a "Test alert snooze" and analyze its characteristics. language of the presentation: Japanese 発表題目: 継続的インテグレーションにおけるテスト連続失敗（Test alert snooze）の実証分析 発表概要: 継続的インテグレーション（CI）は，ソフトウェア変更の度にテストを自動実行される．これにより，常にリリースを維持し，ユーザーに提供するまでの間隔（リリースサイクル）を短縮できる．テストは常に成功が望ましいが，実際の開発では複数回失敗していることが多い．CIでのテストが失敗している状態で，他の不具合を混入させた場合，その混入した不具合が失敗しているテストに紛れてしまう．よって，後に混入した不具合の発見が遅れ，修正するまでの時間がかかる．本研究では，不具合の修正作業をより効率化するために，2回以上連続で失敗しているテストをTest alert snoozeと定義し，その特徴を分析する． | |||||||
| JU YUAN | M, 1回目発表 | ネットワークシステム学 | 岡田 実 | 林 優一 | 東野 武史 | Dipanita Chakraborty | |
|
title: Study on Coupling Optimization Strategies for DWPT Systems
abstract: The increasing adoption of electric vehicles (EVs) and automated guided vehicles (AGVs) has intensified the need for efficient, continuous power-supply solutions. Dynamic wireless power transfer (DWPT) provides a safer and more economical alternative for long-distance and fixed-route charging. This research focuses on improving DWPT system reliability and reducing deployment cost through coupling-structure optimization and the use of economic components. A passive third winding inserted between two adjacent primary coils is proposed to enhance coupling efficiency by reinforcing magnetic interaction without increasing system complexity. In addition, the study analyzes the impact of different ferrite configurations on coupling efficiency and shows that minimizing the use of brittle and expensive ferrite cores further lowers system cost while maintaining effective power transfer. language of the presentation: English | |||||||
| LI ZHIDONG | M, 1回目発表 | ネットワークシステム学 | 岡田 実 | 林 優一 | 東野 武史 | Dipanita Chakraborty | |
|
title: Feasibility Study on Large-scale Hybrid Optical-RF satellites communicaitions
abstract: To overcome severe Ground-to-Satellite Link channel degradation, this study evaluates a massive spatial diversity architecture where numerous LEO satellites relay a ground-based RF signal to a central hub via optical Inter-Satellite Links for coherent combining. An end-to-end MATLAB simulation was developed, modeling key physical layer impairments including atmospheric fading, pointing errors, and multiple optical noise sources such as thermal, shot, Relative Intensity Noise, and signal-signal beat noise. We compared the performance of direct detection, Erbium-Doped Fiber Amplifier (EDFA) pre-amplified and coherent optical receivers. Results demonstrate that while direct detection is insufficient, advanced optical techniques like EDFA pre-amplification and coherent detection are crucial for achieving the high SNR required for reliable communication. language of the presentation: English | |||||||
| 水津 徹久 | M, 1回目発表 | ヒューマンＡＩインタラクション | Sakriani Sakti | 渡辺 太郎 | 大内 啓樹 | Faisal Mehmood | Bagus Tris Atmaja |
|
title: Automatic Generation of a Geospatial Reasoning Benchmark Dataset Integrating Semantic Knowledge and Spatial Constraints
abstract: In recent years, research on the geospatial reasoning ability of large language models (LLMs) has advanced, yet many models have been reported to lack sufficient geographical knowledge and to exhibit limited accuracy and consistency in reasoning. This study represents an initial step toward improving the geospatial reasoning ability of current LLMs by systematically evaluating and analyzing their reasoning performance. Specifically, we automatically generate question–answer pairs that require integrated reasoning over semantic knowledge and spatial constraints by combining a geospatial database (OpenStreetMap) with a knowledge base (Wikidata), and obtain answers from multiple LLMs. The geospatial information involves spatial relations such as distance and containment, while the semantic knowledge refers to attributes of places and their relationships with other entities. Through this evaluation, we clarify both the limitations and the future potential of current LLMs in performing geospatial reasoning. language of the presentation: Japanese 発表題目: 意味と空間の複合知識を要する地理的推論ベンチマークの自動生成 発表概要: 近年，大規模言語モデル（LLM）の地理的推論に関する研究が進んでいるが，多くのモデルについて，地理的知識の不足や推論の正確性・一貫性の欠如が指摘されている． 本研究は，LLMの地理的推論能力の向上に向けた初期的な試みとして，現行モデルの推論能力を体系的に評価・分析することを目的とする．具体的には，地理データベース（OpenStreetMap）と知識ベース（Wikidata）を組み合わせ，地理空間情報および意味的知識の統合的な推論を必要とする質問応答ペアを自動生成し，複数のLLMに対して解答を求めた．地理空間情報は距離や包含関係などの空間的関係，意味的知識は場所の属性や他のエンティティとの関係を指す．本評価を通じて，現行LLMの地理的推論能力の限界と今後の可能性を明らかにする． | |||||||
| 和田 純弥 | M, 1回目発表 | ヒューマンロボティクス | 和田 隆広 | 池田 和司 | 劉 海龍 | 織田 泰彰 | 本司 澄空 |
|
title: Study on a Driving Support Method Using Hanger Reflection Stimuli for
Drivers with Visual Field Deficiencies
abstract: This study investigated a haptic-based driving assistance method using the hanger reflex for d> field defects. The assistance method aimed for in this study is a “self-selecting” driving support system where the hanger reflex device causes a slight head rotation before collision risk increases, but the force is weak, so the human ultimately decides whether to turn toward the target direction. First, as a “basic experiment,” we confirmed reaction times, head movements, and gaze behavior when presenting left/right hanger reflex stimuli under non-driving conditions with no pedestrians present. Next, as a “driving assistance experiment,” we compared stimulus presence versus absence during driving with and without crossing pedestrians to verify the effectiveness of the hanger reflex stimulus device in how drivers respond to hazards. language of the presentation: Japanese 発表題目: 視野欠損ドライバに向けたハンガー反射刺激を用いた運転支援方法の検討 発表概要: 本研究では，部分視野欠損ドライバに向けて，ハンガー反射を用いた触覚ベースの運転支援方法の検討を行った．本研究で目指す支援方法は，衝突リスクが高まる前にハンガー反射装置により頭部をわずかに回転させるが，力は弱いため最終的に対象方向を向くかどうかは人間が判断する「自分で選ぶ」運転支援である．今回は，まず「基礎実験」として，非運転時・歩行者なし条件において左右方向のハンガー反射刺激を与えた際の反応時間，頭部動き，視線挙動を確認した．次に「運転支援実験」として，運転時に横断歩行者あり・なし条件で刺激有無を比較することで，ドライバはハザードに対してどう対応するかハンガー反射刺激装置の効果を確認した． | |||||||
| 竹田 圭汰 | M, 1回目発表 | ユビキタスコンピューティングシステム | 安本 慶一 | 岡田 実 | 諏訪 博彦 | 佐々木 航 | |
|
title: Proposal and Evaluation of a Hierarchical Routing Method for Realizing Large-Scale Satellite Formation Flying
abstract: Recently, dense planar formations of several thousand to tens of thousands of ultra-small satellites have been studied as virtual large antennas to enable high-speed direct communication with general-purpose terminals and expand non-terrestrial networks (NTNs). However, such systems require massive control signaling to all satellites. In this study, we consider about 10,000 cube-shaped satellites with an edge length of 10 cm arranged in a disk with 10 cm spacing, and propose a hierarchical subgroup routing scheme consisting of an anchor satellite, cluster-head (CH) satellites, and child satellites. The routes between CH and child satellites are optimized using the A* algorithm and the 2-Opt method and further shortened by sub-cluster partitioning. As a result, assuming millimeter-wave multi-hop communication, we show that control information can be disseminated to the entire constellation within at most 118 hops and 1.32 ms. language of the presentation: Japanese 発表題目: 大規模衛星フォーメーションフライト実現に向けた階層的ルーティング手法の提案と評価 発表概要: 近年，数千〜数万機規模の超小型衛星を平面上に密集配置し，仮想的大型アンテナとして協調動作させることで，高速ダイレクト通信やNTN拡大が期待されている．しかし，全衛星への制御通信量が膨大となる課題がある．本研究では，約10,000機の10 cm級衛星を10 cm間隔で円盤状に配置し，アンカー衛星・クラスタヘッド（CH）衛星・子衛星からなる階層的サブグループルーティングを提案する．CH–子衛星間経路をA*アルゴリズムと2-Opt法により最適化し，サブクラスタ分割で経路長を抑制した結果，ミリ波マルチホップ通信を仮定した場合でも最大118ホップ・1.32 ms以内で全体の制御情報伝達が可能であることを示した． | |||||||
| 松永 立樹 | M, 1回目発表 | ロボットラーニング | 松原 崇充 | 和田 隆広 | 柴田 一騎 | 鶴峯 義久 | 佐々木 光 |
|
title: Sim-to-Real Transfer of Hook-and-Open Motions Leveraging Force Information
abstract: Hook-based manipulation of objects such as drawers, knobs, and lids is a prerequisite skill for the deployment of robots in domestic environments. Given the diversity of target objects, robots must be capable of adapting to a wide range of hooking scenarios. As these operations are often performed under conditions where visual information is occluded—with fingertips or tools hidden by the object itself—precise control based on force feedback is indispensable. However, dense contact information obtained from force or tactile sensors is highly susceptible to modeling errors in physical simulations, resulting in a significant Sim-to-Real gap. Consequently, directly deploying control policies trained in simulation, which utilize force-tactile data as input, onto real hardware remains a challenging endeavor. This study aims to overcome the challenges associated with the Sim-to-Real transfer of force data and to achieve the robust acquisition of hook-and-open manipulation skills. language of the presentation: Japanese 発表題目フォース情報を活用したsim2realによる引っ掛け開け動作の獲得 発表概覫 家庭環境に存在する引き出し、ノブ、蓋などの引っ掛け操作は、ロボットが過程環境に進出する上で必須のスキルであり、対象物体の多さから様々な引っ掛け操作への対応が求められる。この操作は、指先やツールが障害物に隠れて視覚情報が制限される状況で行われるため、様々な引っ掛け操作に対応するためにも力覚フィードバックに基づく精密な制御が不可欠である。しかし、力覚センサーや触覚センサーから得られる密な接触情報は、センサーに作用する要因が多く物理シミュレーションのモデル化誤差の影響を受けやすいためSim-to-Realギャップが非常に大きい。その結果、シミュレーションで学習した力触覚を入力とする制御ポリシーを実機でそのまま実行することが困難である。本研究は、この力覚データのSim2Realの困難を克服し、引っ掛けて何かを開けるという動作の習得を目指す。 | |||||||
| 越間 龍之介 | M, 1回目発表 | 光メディアインタフェース | 向川 康博 | 和田 隆広 | 藤村 友貴 | 北野 和哉 | |
|
title: Spectral Analysis for Tomb Murals Using a Foundation Model Approach
abstract: This study aims to apply a foundation model to spectral images of tomb murals to achieve digital restoration of damaged sections. Previously, analysis of tomb murals has been conducted using physics-based methods or autoencoders. However, this approach faced challenges due to the scarcity of training data inherent to cultural heritage data. This research aims to achieve more accurate and natural restoration of damaged areas by leveraging the general visual knowledge acquired by the foundation model through extensive pre-training and the detailed color information contained in spectral images. language of the presentation: Japanese 発表題目: 古墳壁画における基盤モデル適用による分光壁画解析 発表概要: 本研究は，基盤モデルを古墳壁画の分光画像に適用し，損傷部分のデジタル修復を実現することを目指す．従来，古墳壁画の解析は，物理モデルベースの手法やオートエンコーダを用いて行われてきた．しかし，文化財特有のデータ希少性による学習データの不足という課題を抱えていた．本研究では，基盤モデルが大規模な事前学習を通じて獲得した汎用的な視覚知識と分光画像が持つ詳細な色彩情報から，損傷部分をより高精度かつ自然に復元することを目的とする． | |||||||
| LI ZHUO | M, 1回目発表 | 大規模システム管理 | 笠原 正治 | 井上 美智子 | 原 崇徳 | ||
|
title: Exploratory Analysis of Structural and Temporal Features for Supply Chain Network Optimization
abstract: Supply chain networks naturally exhibit graph-structured properties. This poster conducts an exploratory analysis of a dataset related to supply chain networks. Prior work on SupplyGraph has demonstrated that incorporating graph structure into the modeling process can improve the understanding and optimization of supply chain behaviors. Building on these existing insights, this study examines the potential value of graph-based modeling approaches for supply chain network optimization tasks. SupplyGraph represents products as nodes and constructs edges through relations such as shared manufacturing plants and product-group associations, enabling the use of graph-based analytical methods. Incorporating structural information can support tasks such as capturing inter-product dependencies, improving forecasting performance, and extending various supply chain analysis capabilities. Based on these observations, this study aims to deepen the understanding of the dataset’s structural characteristics and establish a conceptual and modeling foundation for future development of graph-based temporal models. language of the presentation: English | |||||||
| 寺本 菜々花 | M, 1回目発表 | 情報セキュリティ工学 | 林 優一 | 安本 慶一 | 岡田 実 | 藤本 大介 | |
|
title: Study on the Impact of Process Scaling on Fault Injection Attacks
abstract: Physical attacks against cryptographic modules, particularly fault injection attacks that intentionally induce faults, pose a serious threat. Since fault injection attacks introduce disturbances to cause device failures, fault tolerance depends on device characteristics. Previous research has focused on external signal variations, such as clock rise times, but has not addressed the characteristics of internal integrated circuits (ICs). IC electrical properties change due to process scaling. This research aims to quantitatively evaluate the impact of this process scaling on the success rate of fault injection attacks. Specifically, the same cryptographic algorithm is implemented on three generations of FPGA boards (130nm, 65nm, 20nm), and fault injection attacks are performed to measure their success rates. This evaluation will clarify the effect of scaling on fault injection attacks. language of the presentation: Japanese | |||||||
| 谷崎 文那 | M, 1回目発表 | 情報基盤システム学 | 藤川 和利 | 門林 雄基 | 林 優一 | 新井 イスマイル | |
|
title: The current state of BGP security and RPKI
abstract: Today, various types of communication are conducted over the Internet. The Internet is not managed by a single organization; instead, it is composed of numerous connected networks, known as Autonomous Systems (AS), each managed by a single set of administrative rules. Communication across ASes is carried out based on a protocol called the Border Gateway Protocol (BGP). Since BGP was originally designed under the assumption of a trustworthy environment, such as research institutions, the protocol itself lacks security mechanisms like authenticationand validation. However, because critical internet communications, such as banking transactions involving monetary value, now rely on BGP, ensuring BGP security has become a vital issue. In this research, with the objective of "the healthy development of the Internet," we investigated the security problems inherent in BGP and the approaches being taken to secure BGP. Additionally, we conducted research on the Resource Public Key Infrastructure (RPKI), a BGP security technology currently gaining widespread adoption. language of the presentation:Japanese 発表題目:BGPセキュリティとRPKIの現状 発表概要:今日，様々な内容の通信がインターネットを介して行われている．インターネットは単一の組織が管理しているわけではなく，単一の管理規則によって管理されているネットワークであるAutonomous System （AS）が多数存在し，それらが接続することで成り立っている．ASを跨ぐ通信はBorder Gataway Protocol （BGP）という通信プロトコルに基づいて行われている．BGPは元々研究機関などの信頼できる環境を前提として設計されたプロトコルであるため，BGPそのものには認証・検証などのセキュリティの仕組みが存在しない．しかし，現在では銀行取引など金銭が関係する重要なインターネット上の通信もBGPに依存するため，BGPのセキュリティを確保することは重要な課題である． 本研究においては，「インターネットの健全な発展」を目的とし，BGPが抱えているセキュリティ的な問題やBGPのセキュリティ確保のためのアプローチなどに対して調査を行った．また，現在BGPセキュリティ技術として普及が進んでいるResource Public Key Infrastructure （RPKI）についても調査を行った． | |||||||
| 尾﨑 翠 | M, 1回目発表 | 数理情報学 | 池田 和司 | 清川 清 | 久保 孝富 | 日永田 智絵 | LI YUZHE |
|
title: Gait Disorder Screening for Animals Using DeepLabCut and CEBRA
abstract: Gait disorders in cats can significantly impair their quality of life (QOL), and early detection is important for improving prognosis. However, early-stage gait abnormalities are often difficult for owners to recognize, and natural gait is hard to capture in clinical settings. In addition, conventional video-based analysis requires extensive pose annotation, and its high cost has hindered practical implementation. In this study, we aimed to develop an annotation-free screening approach for detecting cat gait abnormalities by combining DeepLabCut (DLC), a markerless pose estimation method, with CEBRA, a self-supervised learning framework. Using a pretrained DLC model for quadrupeds, we estimated poses from video recordings and obtained pose sequences for each individual. These sequences were then input into CEBRA, which embedded them into a three-dimensional latent space using individual identity and temporal information as auxiliary variables to visualize latent structure in gait patterns. As a result, DLC stably estimated major body points across diverse cases, including both normal and abnormal gait, without the need for manual annotation. In the CEBRA latent space, normal and abnormal gait patterns occupied distinct regions, although some overlap remained near the boundaries. These findings suggest that cat gait features can be extracted without pose annotation and that the latent structure may enable discrimination between normal and abnormal gait. language of the presentation: Japanese 発表題目: DeepLabCutとCEBRAを用いた動物の歩行障害スクリーニング 発表概要: ネコの歩行障害はQOLを損なうため、早期発見は予後改善に重要である。 しかし軽度段階では発見が難しく、診察室では自然な歩行が捉えにくい。また、従来の動画解析手法は膨大な姿勢アノテーションを必要とし、そのコストの高さから実用化が進んでいない。本研究では、マーカーレス姿勢推定手法の DeepLabCut（DLC）と自己教師あり学習手法の CEBRA を組み合わせ、姿勢アノテーションを用いないネコの歩行異常スクリーニング手法の構築を目指した。まず、DLCの四足動物用事前学習済みモデルを用いて動画から姿勢を推定し、姿勢系列を取得した。次に、この姿勢系列をCEBRAに入力し、個体IDと時間情報を補助変数として3次元潜在空間に埋め込むことで、歩行パターンの潜在的構造を可視化した。その結果、DLCはアノテーションなしでも正常例・異常例を含む多様な症例において主要部位を安定して推定し、CEBRAの潜在空間では正常歩行と異常歩行が異なる領域に偏って分布することが確認された。これらの結果から、姿勢アノテーションを行わずにネコの歩行特徴を抽出し、歩行の正常・異常を区別できる可能性が示された。 | |||||||
| 衛 飛那汰 | M, 1回目発表 | 生体画像知能 | 大竹 義人 | 池田 和司 | Soufi Mazen | Gu Yi | |
|
title:Phase Normalization for a 4DCT-Based AI System for Swallowing Treatment Support
abstract: The evaluation of dysphagia, which is increasing with the aging population, is currently limited to 2D observations using videofluorography or endoscopy. Consequently, it is impossible to grasp swallowing functions in three dimensions quantitatively. Furthermore, diagnoses rely heavily on the physician's subjective evaluation, making patient-specific analysis difficult. This study aims to establish a quantitative, patient-specific evaluation method by using 4D-CT images to capture the dynamics of the tongue and surrounding tissues. Currently, we have performed phase recognition for some cases and visualized shape changes. In the future, we plan to proceed with data interpolation using non-rigid registration and work on creating an average swallowing motion model and a phase recognition algorithm. language of the presentation: Japanese 発表題目: 嚥下治療支援を目的とした4DCT解析用AIのための時相正規化 発表概要: 高齢化に伴い増加する嚥下障害の評価は、嚥下造影や内視鏡による2次元観察に留まり、嚥下機能の3次元的かつ定量的な把握はできない。また、診断は医師の主観的評価によるところが大きく、患者個別の解析が困難である。本研究では、4DCT画像を用い、舌や周辺組織の動態を捉え、患者個別の定量的評価手法の実現を目指す。現在、一部の症例について時相認識を行い、形状変化の可視化を行った。今後は、非剛体レジストレーションを用いたデータ補完作業を進め、嚥下の平均的な動作形状モデルと時相認識アルゴリズムの作成に取り組む予定である。 | |||||||
| 小泉 孝太朗 | M, 1回目発表 | 脳・行動モデリング(計算神経科学) | 田中 沙織☆ | 川鍋 一晃 | 杉本 徳和 | 荻島 大凱 | |
|
title: Confidence Dynamically Controls Abstract Learning
abstract: Humans can learn from small samples efficently even in unknown enviorment. Abstraction extract relevant imformation is important to achive this awesome ability. In the previous research, it was shown that humans use abstraction.Not only the previous research but also other researches indicates confidence which is the degree how well you can perform well may play an key role to realize abstraction. However, the computational principle of this remains unknown. The suggested model would be presentated in this presentation. language of the presentation: Japanese 発表題目: 自信の動的な抽象学習制御 発表概要:ヒトは未知の環境でも小サンプルから効率的に学習することができる。この際に関連がある情報のみ取り出す抽象化が重要となる。先行研究では、ヒトは抽象化を用いて効率的に学習していることが示された。先行研究だけでなく他の研究でも自信が抽象化を実現するのに重要だと示唆されている。しかし、その計算原理はいまだによくわかっていない。今回の発表ではモデルの仮説を提案する。 | |||||||
| 平川 稜真 | M, 1回目発表 | 自然言語処理学 | 渡辺 太郎 | 荒牧 英治 | 上垣外 英剛 | 坂井 優介 | |
|
title: Analysis of Internal Mechanisms of Dialect Generation in Large Language Models
abstract: This study aims to analyze how large language models (LLMs) internally represent and generate variations in speech styles such as dialects. Focusing on Kansai and Tohoku dialects, we visualize intermediate layer outputs using the Logit Lens technique to identify layers responsible for dialectal endings and vocabulary. By statistically analyzing evaluation metrics based on these findings, we quantitatively evaluate how dialectal features emerge and evolve during the generation process. This research contributes to improving the transparency and reliability of language style control in LLMs. language of the presentation: Japanese 発表題目: 大規模言語モデルにおける方言生成過程の内部機序解析 発表概要: 本研究は，大規模言語モデル（LLM）が方言などの話し方の違いをどのように内部表現し，生成しているかについて解析することを目的とする．特に関西方言と東北方言を対象にモデルの各層出力に対してLogit Lensにより可視化を行い，語尾や語彙が出現する層を特定し統計的に評価指標を解析することで，生成過程における方言特徴の変化を定量的に評価する．本研究は，言語スタイル制御の透明性・信頼性向上に寄与することを目指す． | |||||||
| LU TONG | M, 1回目発表 | 計算システムズ生物学 | 金谷 重彦 | 松本 健一 | MD.Altaf-Ul-Amin | ||
|
title: Multimodal Health Data-Based Adverse Drug Effect Management System for Type 2 Diabetes
abstract: The core objective of this project is to construct an intelligent health monitoring system based on Graph Convolutional Neural Network (GCN) technology, achieving a complete workflow from health information collection and intelligent analysis to medical advice generation, ultimately accomplishing causal relationship analysis between adverse drug reactions and physiological information. The project utilizes smart wearable devices to collect patients' motion data, and employs Graph Convolutional Neural Network models to identify fifteen types of daily activities across six major categories, including reading, watching TV, drinking water, taking medication, lying down, sleeping, sitting, washing hands, washing face, excretion, jogging, walking, exercising, eating, and other behaviors. language of the presentation: English | |||||||