コロキアムB発表

日時: 02月24日 (Tue) 2限目(11:00 - 12:30)


会場: L1

司会: 佐々木 光
LOBASKIN EGOR D, 中間発表 数理情報学 池田 和司☆, 川鍋 一晃, 杉本 徳和, 田中 沙織
title: Metric learning of Trans-diagnostic Biomarkers from Human Brain Functional Connectivity
abstract: Advances in machine learning have transformed data-rich clinical domains such as genomics and radiology, yet psychiatry lags in the clinical adoption of computational methods. Neuroimaging in mental health faces three major structural constraints: severe data scarcity relative to dimensionality, cross-site heterogeneity in image acquisition, and ethical concerns around deploying opaque AI in sensitive clinical contexts. We propose a geometry-aware framework that replaces reductive classification with metric learning in the spectral domain of functional connectivity matrices, mapping patients into a continuous transdiagnostic landscape that respects symptomatic overlaps between conditions as well as individual variation. Leveraging Fourier analysis on the graph Laplacian, we learn spectral filters that isolate pathology-relevant network modes, while adversarial regularization ensures embeddings remain grounded in clinical plausibility. Critically, we uncover a persistent confound: site and protocol biases survive both geometric alignment and regression-based harmonization, manifesting in eigenvector orientations. This reveals a blind spot in current practices and calls for spectral-aware approaches to domain adaptation.
language of the presentation: English
 
児玉 創太郎 M, 2回目発表 脳・行動モデリング 田中 沙織, 池田 和司, 久保 孝富, 荻島 大凱
title: Mathematical Modeling of Intolerance of Uncertainty in Anxiety Disorders
Anxiety disorders are prevalent psychiatric conditions characterized by excessive fear and anxiety that impair daily functioning. Although cognitive behavioral therapy (CBT), particularly exposure therapy, is effective, many patients show limited response or relapse. This suggests the need to move beyond a framework focused solely on fear and avoidance of explicit threats. Individuals with anxiety disorders experience heightened distress not only in the presence of actual threat, but also in uncertain situations. This tendency is captured by intolerance of uncertainty (IU), defined as perceiving uncertainty as threatening and unbearable, leading to avoidance and safety behaviors. Uncertainty can be divided into risk (known probabilities) and ambiguity (unknown probability structure). Prior research links risk to attentional bias toward threat and ambiguity to biases in estimating environmental volatility. However, these accounts do not explain why uncertainty itself becomes intolerable. To address this, we modeled participant behavior in an online interview task using a variational Bayesian framework based on a partially observable Markov decision process (POMDP). Participants inferred evaluation tendencies from facial expressions and feedback while deciding whom to appeal to. The model estimates biases in observational noise, environmental volatility, and threat-related valuation. We hypothesize that heightened anxiety leads to avoidance of negative information, impaired differentiation between change and noise, and persistent uncertainty, ultimately reinforcing intolerance of uncertainty.
 
佐野 海士 D, 中間発表 脳・行動モデリング 田中 沙織, 池田 和司, CAI LIN, 荻島 大凱
Colloquium B - 2421033
Title Beliefs Under Uncertainty: Risk Representation and Source Trust in Societal Issues
Abstract

Effective responses to societal hazards often require collective coordination, yet misinformation can foster non-cooperative beliefs that undermine risk communication. This talk presents my doctoral work on how people form and update beliefs under uncertainty, combining two complementary studies.

Study 1 examines the representation of societal hazards within the psychometric paradigm. Using a Japanese translation of the classic risk-perception items, I map hazards in a two-dimensional space of Dread and Unknown and validate the measurement model for individual-level scoring (including cross-loadings and excluding an ambiguous "not scientific" item). Reliability is evaluated with two approaches to internal consistency and with test-retest stability across two sessions.

Study 2 (planned) targets the updating mechanism under realistic conditions where ground truth is rarely clear and feedback is noisy. Building on source-learning and repetition-interpretation research, I propose an experiment that manipulates feedback quality (full vs. noisy) and explanatory narratives (positive vs. negative vs. none), including doubt-scenario types that may induce categorical source labels. The central hypothesis is that explanation-driven labeling will bias trust updating more strongly under noisy feedback, helping clarify when non-cooperative beliefs emerge and persist, and informing more effective risk and science communication.

 

日時: 02月24日 (Tue) 2限目(11:00 - 12:30)


会場: L2

司会: 笹田 大翔
WEI XUEFENG D, 中間発表 自然言語処理学 渡辺 太郎, Sakriani Sakti, 上垣外 英剛, 坂井 優介
title: Toward Fair and Interpretable Health Advice: Auditing and Mitigating Nutritional Bias in LLM-Based Meal Recommendations
abstract: Large language models are increasingly used for personalized health advice, yet meal recommendations can vary systematically with demographic cues, raising health-equity concerns. My goal is to develop a closed-loop framework to measure, explain, and mitigate nutritional bias in LLM-based meal recommendation: I will introduce nutrition-grounded benchmarks and disparity metrics, analyze how demographic cues shape intermediate semantic choices, and design controllable interventions evaluated in multi-turn and cross-cultural settings. A completed milestone is an audit study using a menu-constrained evaluation pipeline that reveals persistent race- and gender-associated disparities.
language of the presentation: English
 
ZHOU XUAN D, 中間発表 自然言語処理学 渡辺 太郎, Sakriani Sakti, 上垣外 英剛, 坂井 優介
title: AutoPilot GeoAgents: A Multi-Agent, Tool-Managing System for Fully Automated Multimodal Remote Sensing Analysis
abstract: AutoPilot GeoAgents is a fully automated, multi-agent system for multimodal remote sensing analysis. Given a user query, it decomposes the task, manages external tool calls, verifies intermediate results, and iterates in a closed loop until producing a final, actionable report. The research plan is organized into two subgoals: (1) establishing core tool-using capability and benchmarked evaluation for multimodal RS understanding, (2) developing a planner–executor–verifier–reporter multi-agent loop for reliable end-to-end task completion without human intervention.
language of the presentation: English
 
YU SHUYI M, 2回目発表 自然言語処理学 渡辺 太郎, Sakriani Sakti, 上垣外 英剛, 坂井 優介
title: Personalized Entity Familiarity Prediction via Popularity-guided Directed Label Propagation on Knowledge Graphs
abstract: To bridge the "knowledge gap" in specialized texts, we propose a framework that predicts personalized entity familiarity by leveraging Knowledge Graphs and Graph-based Active Learning. Our system dynamically constructs subgraphs using Wikipedia2Vec embeddings and YAGO 4.5 relations , then identifies the most informative entities via Sigma-optimality for user labeling. By implementing Label Propagation across these directed semantic flows , we accurately map a user's knowledge boundaries to automatically generate personalized glossaries.
language of the presentation: English