コロキアムB発表

日時: 06月13日 (Fri) 3限目(13:30-15:00)


会場: L1

司会: Perusquia Hernandez Monica
CAO SHIYI M, 2回目発表 数理情報学(計算神経科学) 池田 和司☆, 川鍋 一晃, 杉本 徳和, 田中 沙織
title: Adaptive distortions of confidence under asymmetric environments and internal constraints
abstract: Confidence judgments play a critical role in adaptive cognition, yet they are typically modeled as static functions of internal evidence. Here, we show that metacognitive sensitivity—the ability to discriminate between correct and incorrect decisions—emerges from dynamic interactions among decision criteria, stimulus sampling, and evidence variance. Using simulations based on signal detection theory (SDT), we demonstrate that environmental asymmetries distort metacognitive resolution (AUROC2) through two key mechanisms: shifts in sampling priors and systematic variation in evidence variance. Crucially, we find that variance and sampling asymmetries are not independent but exhibit a structured relationship—suggesting that environments with rare stimuli may simultaneously produce noisier evidence, compounding metacognitive distortion. Neural network models trained under such conditions replicate these distortions endogenously, even when overall accuracy is preserved. Further, behavioral data reanalyzed with reaction time as a confidence proxy show parallel distortions linked to both prior probability and decision bias. These findings challenge fixed-evidence models of metacognition and point toward a more general framework in which metacognitive precision reflects the joint statistics of base rates, internal noise, and strategic decision policies.
language of the presentation: English
 
山口 晴久 M, 1回目発表 数理情報学(計算神経科学) 池田 和司☆, 作村 諭一, 川鍋 一晃, 田中 沙織
 
吉岡 春彦 M, 2回目発表 ソフトウェア工学 松本 健一, 池田 和司, Raula Gaikovina Kula, 嶋利 一真
title: Proposing a Distributed Representation of Eye Movements for Program Comprehension
abstract: Program comprehension is a critical activity in software development. Eye tracking has been employed to capture this process. However, traditional analysis methods suffer from a lack of objectivity and efficiency, as they often rely on manual labeling of fixated elements and single-faceted evaluations. In this study, we propose eye2vec, a method to automatically map gaze data to source code elements—including words, syntax, and their relationships—and convert them into high-dimensional distributed representations. Through an experiment involving a bug detection task in Java, we demonstrate that (1) the resulting vectors can effectively distinguish between developers based on their bug-finding performance, and (2) our method can extract source code elements that are instrumental in bug detection.
language of the presentation: Japanese
発表題目: プログラム理解における視線移動の分散表現の提案
発表概要: プログラムの理解はソフトウェア開発において重要な活動である.その過程を捉える手段として視線計測を用いた分析がこれまでに行われてきた. しかし,従来の分析手法は注視対象の手動ラベリングや単一観点での評価に依存しており,客観性と分析効率に課題がある. 本研究では,視線とソースコードの各要素である単語や構文,単語間の関係性を自動的に対応づけ,高次元ベクトルに変換する手法eye2vecを提案する. Javaのバグ発見タスクへと適用した評価実験の結果,(1) ベクトルがバグ発見の有無によって開発者ごとに分離されること,(2) バグ発見に寄与するソースコード要素を抽出可能であることを示した.
 
LERTBANJONGNGAM SILA M, 2回目発表 ソフトウェア工学 松本 健一, 安本 慶一, Raula Gaikovina Kula, 嶋利 一真, Fan Youmei
title: From Vision to Code: Evaluating Large Multimodal Models for Code Generation in Notebooks
abstract: Large Multimodal Models (LMMs) show potential in generating code from visual inputs, but their effectiveness in real-world notebook environments is still unclear. This study evaluates how well LMMs generate executable code from visualization images in Jupyter notebooks. Results show that while models follow common practices (e.g., using matplotlib.pyplot and seaborn), only about 50% of the generated code is executable. Frequency and distribution graphs perform best. Most failures are due to missing context—especially the lack of actual data—leading to code mistakes and logic bugs. Overall, LMMs are promising but still face key limitations in reliable code generation from images alone.
language of the presentation: English
 
菊池 尊勝 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, 松井 智一
title: Cross-modal Daily Activity Recognition based on Fixed Sensor and Digital Twin
abstract: In recent years, privacy concerns have driven growing interest in human activity recognition (HAR) methods that rely solely on Fixed sensors, avoiding the use of cameras or microphones. Multimodal learning approaches that leverage contrastive learning to integrate Fixed sensors with other modalities have shown promise. However, several challenges remain, including the limited informativeness of Fixed sensors, poor generalization across different households due to variations in sensor placement and layout, and the high cost of collecting real-world data. This study addresses these challenges through three key strategies: (1) enhancing feature representations by semantically integrating sensor data with other modalities, (2) improving generalization by modeling household-specific characteristics, and (3) introducing an indoor digital twin framework that enables virtual data augmentation and simulation. These efforts aim to realize practical and privacy-preserving HAR using only Fixed sensors.
language of the presentation: Japanese
タイトル:宅内デジタルツインを活用した環境設置型センサによる生活行動認識手法の検討
アブストラクト:近年、プライバシー保護の観点から、カメラやマイクを用いずに環境設置型センサのみで生活行動を認識する研究が注目されている。特に、対照学習により環境設置型センサと他モーダルを組み合わせたマルチモーダル学習は有望であるが、環境設置型センサ固有の制約や家庭間差、データ収集コストなどが課題である。本研究では、(1)センサ情報の意味的統合による特徴表現の強化、(2)家庭特性のモデリングによる汎化性の向上、(3)仮想空間上でのデータ拡張を可能にする宅内デジタルツインの設計を通じて、環境設置型センサによる実用的な行動認識の実現を目指す。
 

日時: 06月13日 (Fri) 3限目(13:30-15:00)


会場: L2

司会: Chakraborty Dipanita
SHAHZADI ANAM M, 2回目発表 光メディアインタフェース 向川 康博, 松原 崇充, 舩冨 卓哉, 藤村 友貴, 北野 和哉
title:ShuttleHitFormer: A Multi-Feature Transformer Model for Shuttle Hit Detection in Badminton
abstract: Hit detection in badminton is critical for automated match analysis but remains challenging due to the shuttle's high-speed motion, frequent occlusions, and complex player-shuttle interactions. We propose ShuttleHitFormer, a novel Transformer-based architecture that integrates multi-stream encoders (shuttle trajectories, player poses, court-net keypoints), hierarchical cross-attention fusion, and a temporal Transformer to model spatial-temporal dependencies. Our method demonstrates strong performance on the ShuttleSet22 dataset, significantly outperforming existing approaches. Ablation studies further validate the effectiveness of multi-feature fusion. This work advances sports analytics by showing that Transformer architectures, when tailored to domain-specific challenges, can deliver fine-grained tactical insights in fast-paced racket sports.
language of the presentation: English
 
SYAFRUDIN RAIS AKHDAN M, 2回目発表 ヒューマンロボティクス 和田 隆広, 松原 崇充, 織田 泰彰, 劉 海龍, 本司 澄空
title: Human Guidance Along a Defined Path Using a Robotic Dog with Harness Interaction Based on Tractor-Trailer and Data-Driven Models
abstract: Robot guide dogs offer a promising alternative to real guide dogs, whose availability remains limited, in supporting the safety and mobility of blind or visually impaired (BVI) individuals. For such systems to be effective, they must accurately monitor human motion and provide real-time corrective feedback an inherently complex task due to the requirement of precise robot control while accounting for human variability. This research focuses on improving guidance precision by incorporating the human into the control loop using a tractor-trailer model. The system continuously estimates the human’s position and pose to correct trajectory deviations, with performance evaluated in simulated obstacle avoidance scenarios. Results show effective lateral error correction. However, discrepancies persist between the expected human motion and the actual measured responses, largely due to the dynamic and unpredictable nature of human behavior. To address this limitation, a data-driven model is proposed to better capture and predict human motion, thereby enhancing the system's adaptability and accuracy.
language of the presentation: English
 
NGUYEN NGOC HUY M, 2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, 郭 政佑
title: Trajectory prediction with discriminative feature embedding for catching diverse and unseen objects
abstract: Robotic catching of flying objects demands accurate future trajectory prediction from early-stage due to limited flight time. Some objects exhibit complex aerodynamics, making it difficult to predict their future trajectories from short historical trajectories. Previous studies use recurrent neural networks to predict the future trajectories of objects with complex dynamics. However, early-stage trajectories often appear similar across multiple objects, making existing methods fail to differentiate between trajectories of different objects. Furthermore, when applied to unseen objects, these methods cannot match unseen object trajectories to similar seen trajectory features, resulting in inaccurate predictions. In this study, we propose a trajectory prediction method with discriminative feature embedding for catching diverse and unseen objects. The core idea is to project trajectories into a feature space that captures object-specific dynamics. This embedding enables the model to distinguish between objects from limited historical trajectories and accurately predict future motion for diverse and unseen objects. Experiments on 20 real-world aerial trajectories demonstrate that our method can clearly distinguish trajectory features across multiple objects and map the trajectories of unseen objects to similar seen objects. This capability improves prediction accuracy for diverse and unseen objects, resulting in higher catching success rates.
language of the presentation: English
 
DANG PHUONG NAM M, 2回目発表 ヒューマンロボティクス 和田 隆広, 清川 清, 織田 泰彰, 劉 海龍, 本司 澄空
title: Passenger-Centered Voice Interfaces: Enhancing Autonomous Personal Mobility Vehicles with Personality-Tuned External Human-machine Interface
abstract: Autonomous Personal Mobility Vehicles (APMVs) are designed to provide efficient short-distance transportation in mixed-traffic environments. Due to their open-body design, APMVs require effective communication with pedestrians through External Human-Machine Interfaces (eHMIs), which directly impacts the passenger experience. To enhance this experience, it is essential to develop personalized voice-based eHMIs that consider both the content and design of the communication. This study investigates how passengers’ preferences for voice-based eHMIs are influenced by their Big Five personality traits, using data collected from on-vehicle user experiments. Based on these insights, we propose a graph-based recommendation model that selects the most suitable voice-based eHMI for each passenger, enabling adaptive and personality-aware interaction in autonomous mobility systems.
language of the presentation: English
 

日時: 06月13日 (Fri) 3限目(14:30-15:00)


会場: L3

司会: 范 優美
ONG HANS JARETT JIM D, 中間発表 数理情報学 池田 和司, 笠原 正治, 久保 孝富, 日永田 智絵, LI YUZHE
title: Advancing Causal Learning: From Structured Data Discovery to Meta-Learning and Representation Learning
abstract:

While machine learning models excel at pattern recognition, they struggle to understand cause-and-effect relationships—limiting their interpretability, robustness, and generalization. My dissertation addresses this through four projects spanning three data settings:

Structured (tabular) data: Two complementary approaches enhance causal discovery: (1) LiNGAM-SPP improvements using pairwise likelihood ratios (replacing kNN-based mutual information), node-skipping for user constraints, and shortest-path distribution analysis to detect latent confounders and predict performance, and (2) Compression-based Dependence Measure (CDM) that approximates Kolmogorov complexity using standard compressors for reliable independence testing in small-sample regimes.

Sparse interventional data: MetaCaDI combines Model-Agnostic Meta-Learning with differentiable DAG sampling to address data scarcity, scalability, and unknown intervention targets. Each intervention context becomes a MAML task, enabling few-shot adaptation to new experimental settings while jointly learning shared causal structure.

Unstructured (image) data: We extend CausalVAE with differentiable DAG sampling to guarantee valid causal graphs during training while generalizing from linear Gaussian to non-linear additive structural causal models. This recovers causally disentangled latent representations that support controlled interventions and counterfactual reasoning.

Together, these projects demonstrate how causally-informed techniques spanning discrete graph search to continuous differentiable learning can enhance ML interpretability, robustness, and generalization across diverse data modalities.


language of the presentation: English