コロキアムB発表

日時: 06月26日 (Thu) 3限目(13:30-15:00)


会場: L1

司会: PHAM HOAI LUAN
安齊 卓哉 D, 中間発表 数理情報学 池田 和司, 松本 健一, 久保 孝富, 田中 沙織
title: Multi-level metacognitive signals to facilitate flexible behaviour
abstract:Humans are adept at making efficient decisions in a volatile environment where the sources of outcomes and perceptual information are uncertain and noisy. For instance, a skilled fisherman may perform well to decide whether to change their tools or locations after catching a few fish, even without direct evidence of a shoal of fish in the ocean. My research investigates how humans form distinct metacognitive signals, how their prior knowledge shapes their perception, and how psychiatric disorders affect their behaviour. These ongoing investigations suggest that humans simultaneously form distinct metacognitive signals, improving credit assignment jointly. Additionally, in a different line of work, I found that humans can infer others’ mental states through perceptual information with prior knowledge.
language of the presentation:Japanese
 
西村 虎太郎ジェームス D, 中間発表 数理情報学 池田 和司, 作村 諭一, 久保 孝富, 日永田 智絵, LI YUZHE
title: Extraction of Material-Condition Optimization Metrics from FSB Charge–Discharge Data Analysis
abstract: In this study, we objectively quantify the charge behavior of fluoride shuttle batteries (FSBs)—a promising class of next-generation secondary batteries—and extract data-driven features that support materials development and experimental-condition optimization. Voltage–capacity curves obtained from charge–discharge cycling tests are processed to derive primary metrics such as capacity retention, Coulombic efficiency, hysteresis area, and dQ/dV peak positions; their correlations with controllable factors including temperature, current density, and active-material composition are then evaluated statistically. Our current efforts focus on feature extraction and visualization, and the insights gained will guide the development of experimental-design methods that explicitly account for model uncertainty. This presentation reports the constructed metric framework, preliminary correlation analysis, and future outlook.
language of the presentation: Japanese
発表題目:FSB充放電データ解析による材料条件最適化指標の抽出
発表概要: 本研究では、次世代二次電池として期待されるフッ化物シャトル電池(FSB)の充電挙動を客観的に定量化し、材料開発と実験条件最適化に資するデータ基盤を構築する。 充放電サイクル試験で取得した電圧–容量曲線を処理し、容量保持率・クーロン効率・ヒステリシス面積・dQ/dVピーク位置などの一次指標を抽出したうえで、温度・電流密度・活物質組成との相関を統計的に評価した。 現在は特徴量抽出と可視化に注力しており、得られた知見を踏まえて今後はモデル不確実性を取り込んだ実験設計手法を検討する予定である。本発表では、これまでに構築した指標体系と基礎的相関解析の成果、および今後の展望を報告する。
 

日時: 06月26日 (Thu) 3限目(13:30-15:00)


会場: L3

司会: 佐々木光
河野 真有香(オンライン) D, 中間発表 サイバネティクス・リアリティ工学 清川 清, 向川 康博, 内山 英昭, Perusquia Hernandez Monica, 平尾 悠太朗
title: Simulating Visual-Linguistic Processing Tendencies of Children with ASD Using Multimodal Large Language Models
abstract: As of 2022, 1 in 100 children worldwide is diagnosed with Autism Spectrum Disorder (ASD). Early detection, intervention, and support are known to contribute to better outcomes for individuals with ASD. Effective intervention and support depend heavily on the involvement of caregivers and supporters surrounding children with ASD. However, understanding how children with ASD perceive objects and situations remains challenging, making it difficult to evaluate the appropriateness of interventions and support strategies. Large Language Models (LLMs) have attracted attention from an engineering perspective for their potential to mimic human behavior. This study aims to simulate the visual-linguistic processing tendencies of children with ASD using Multimodal Large Language Models (MLLMs). I investigate the visual-linguistic processing tendencies of MLLMs focusing on a task and corpus in which children with ASD and typically developing (TD) children create stories based on pictures. In a preliminary experiment, I examined whether MLLMs could accurately identify stories created by children with ASD. The results showed an identification accuracy ranging from 11% to 33%. Building on these initial findings, as well as prior research on LLMs related to Theory of Mind (ToM), language acquisition, and knowledge formation, I have developed a plan to: (1) test the hypothesis that MLLMs exhibit generation tendencies similar to those of children with ASD, and (2) approximate the generation tendencies of MLLMs to those of children with ASD through prompting techniques, thereby simulating their visual-linguistic processing patterns. In the future, I hope to utilize the findings of this study to develop a system that presents the thought processes of children with ASD when they receive visual information. Such a system could help caregivers better understand how children with ASD perceive the world around them. In this presentation, I will report the results and discussion of the preliminary experiment, as well as outline our future experimental plans.
language of the presentation: Japanese
発表題目: マルチモーダル大規模言語モデルによるASD小児の視覚-言語処理傾向シミュレーション
発表概要: 2022年現在,世界の小児の100人に1人が自閉スペクトラム症 (ASD) と診断されている.ASDは早期発見・介入・支援が良好な予後をもたらすと知られている.有効な介入・支援には,ASD小児の周囲の支援者の関わり方が重要だが,ASD小児の事物の捉え方の理解は困難で支援・介入の適切性評価における課題となっている.大規模言語モデル (LLM) は工学的観点から人間の模倣可能性が注目されている.本研究では,マルチモーダル大規模言語モデル (MLLM) によりASD小児の視覚-言語処理傾向をシミュレートすることを目指す.ASD小児と定型発達 (TD) 小児が絵を見て物語を作るタスクとコーパスに着目し,MLLMの視覚-言語処理傾向を調査している.初期実験では,MLLMがASD小児の作った物語を正しく識別可能かを調査したが,11%-33%の識別精度であると分かった.初期実験とLLMの心の理論 (ToM) や言語獲得・知識形成に関する先行研究を踏まえ,(1) MLLMはASD小児に近い生成傾向を示すという仮説の検証,(2) プロンプト技術による生成傾向のASD小児の傾向への近似を行い,ASD小児の視覚-言語処理傾向をシミュレートする計画を立てている.将来的に本研究の成果を用い,ASD小児の視覚情報受容時の思考提示システムを開発すれば,支援者にASD小児の事物の捉え方の理解を促せると期待している.本発表では,初期実験の結果・考察と今後の実験計画について報告する.
 
WEI XIN D, 中間発表 サイバネティクス・リアリティ工学 清川 清, 田中 沙織, 内山 英昭, Perusquia Hernandez Monica, 平尾 悠太朗
title: Unobtrusive Refractive Power Monitoring Using Eyewear-Based EOG and Eye-Tracking
abstract: Early detection of visual impairments remains a persistent challenge, especially due to the subtle and often unnoticed nature of early-stage symptoms. Recent works have attempted to transition clinical tests to home-based services or develop innovative diagnostic methods, but most approaches remain self-initiated and discrete. In this study, we investigates the feasibility of estimating refractive error using biosignals, specifically electrooculography (EOG) and eye-tracking data. We recruited thirty-nine participants used optometry trial lenses to simulate different refractive conditions. Participants performed a series of visual tasks while their eye movements were recorded. Our analysis explored both unimodal (EOG or eye tracking) and multimodal approaches under subject-dependent and subject-independent scenarios. Our goal is to develop an unobtrusive refractive power monitoring system that can track vision status during daily life, thereby to achieve early detection of refractive errors.
language of the presentation: English