이 문서는 NotebookLM 코퍼스(100편 naturalistic fMRI 논문)에 실행한 20개 전문가급 쿼리의 전문과 설계 배경을 제공합니다.
| Category | # Queries | Rationale |
|---|---|---|
| A. Methodological Foundations | 3 | 통계적 추론, 자극 표준화, 동작/아티팩트 — 모든 결론의 신뢰성 토대 |
| B. Neural Representation & Dynamics | 3 | 시간 계층, 공유 vs 개별 부호, 다중 모달 — naturalistic 고유 기여 |
| C. Memory, Events, Prediction | 3 | 사건 분할, 기억 공고화, 예측 — 연속 경험 → 이산 표상 |
| D. AI-Neuro Alignment | 4 | LLM 정합, 기초 모델, 생성적 디코딩, 신경 영감 AI — 가장 빠른 성장 프런티어 |
| E. Individual Differences & Clinical | 3 | 임상 우위 기전, 종단 추적, 지문 해석 — 사회적 임팩트 영역 |
| F. Developmental & Cross-Species | 2 | 발달 궤적, 종간 정합 — naturalistic 특히 강력한 두 영역 |
| G. Frontiers & Meta-Science | 2 | 문화적 편향, 영화 너머 — 분야의 외연과 건전성 |
모든 쿼리는 동일한 구조:
- (a) SOTA: 코퍼스가 보여주는 현재 합의
- (b) Outstanding: 미해결 쟁점
- (c) AI angle: Neuro-AI가 제공하는 지렛대
How do the corpus papers infer neural representations from non-stationary, trial-free naturalistic data?
(a) Dominant frameworks (ISC variants, encoding models, HMM, MVPA, information-theoretic measures) and their underlying assumptions (b) Statistical power limitations, generalizability challenges, multiple-comparison issues under dependent samples (c) AI-driven advances — counterfactual stimulus generation via foundation models, information bottleneck theory, causal representation learning
Synthesize across multiple papers with specific citations.
Why: Naturalistic fMRI는 repeated trial이 없음 → 전통 통계 부적합. ISC/HMM 등 새로운 프레임워크 필요성과 한계.
Answer file: ../answers/Q01_statistical_inference.md
The field leans heavily on a limited canon of stimuli (Sherlock, Forrest Gump, Pixar shorts like Partly Cloudy, Black Mirror episodes).
(a) What evidence exists in the corpus for cross-stimulus generalization vs. stimulus-specific findings? (b) Standardization efforts and shared datasets — Narratives, StudyForrest, Natural Scenes Dataset, HCP 7T movies, CamCAN? (c) Can foundation-model embeddings (CLIP, Whisper, LLMs) enable stimulus-invariant inference — treating any naturalistic content as a sample from a "stimulus basis space"?
Synthesize with citations.
Why: Sherlock-effect (특정 영화가 과도하게 인용됨) 극복 필요. Foundation model embeddings로 stimulus-invariant inference 가능성 탐구.
Naturalistic paradigms inherently maximize head motion (laughter, startle, eye tracking) and physiological variation.
(a) How do corpus papers handle head motion during movie watching, physiological noise correction, engagement-motion coupling? (b) Motion/physiology as pure confound vs. informative signal about engagement/arousal — current debate (c) Deep-learning solutions — motion-aware denoising networks, self-supervised artifact removal, real-time prospective correction
Synthesize with citations.
Why: Naturalistic이 rest 대비 motion 낮지만 여전히 artifact. Motion을 engagement signal로 재해석 가능성.
Answer file: ../answers/Q03_motion_and_artifact_handling.md
Hasson's temporal receptive window hierarchy posits progressively longer integration timescales from sensory to associative cortex.
(a) Corpus evidence for hierarchy universality across stimulus types (narrative/music/silent film); refinements from scale-free dynamics (2025) (b) Hemodynamic confound vs. true neural timescale hierarchy; dissociation methods (c) Can state-space AI models (Mamba, transformers with learned timescale priors) recapitulate hierarchy from brain-agnostic training?
Synthesize with citations.
Why: Temporal hierarchy는 naturalistic fMRI의 signature finding. BOLD 아티팩트 vs 실제 뇌 속성 구분 중요.
The tension between inter-subject correlation (ISC, capturing shared variance) and fingerprinting (capturing idiosyncratic variance) is central in naturalistic fMRI.
(a) Which brain regions and computations are shared vs. idiosyncratic; how this varies by stimulus type, cognitive demand, individual (b) Neurobiological interpretation — hierarchical level (sensory vs. associative), feedforward vs. recurrent, universal vs. culturally-learned (c) How can hyperalignment-based AI embeddings and individual-level foundation models simultaneously capture shared + personal structure?
Synthesize with citations.
Why: 집단 vs 개인의 productive tension. Hyperalignment가 두 측면 동시 포착 가능한지 검증.
Answer file: ../answers/Q05_shared_vs_idiosyncratic_codes.md
Natural stimuli are inherently multimodal (visual + auditory + linguistic + embodied).
(a) Neural binding mechanisms — convergence zones (STS), temporal synchronization, shared embedding spaces (b) Unresolved questions — temporal precedence of modalities, bottom-up vs. top-down binding, role of prediction (c) How do multimodal AI models (CLIP, Flamingo, Gemini, video-language foundation models) predict cross-modal naturalistic activity?
Synthesize with citations.
Why: 2-branch DNN이 early cross-modal 실패 → 뇌는 V1에 acoustic info 표상. 생물학적 타당 아키텍처 필요.
HMM-inferred event boundaries are central in the corpus but their relation to subjective/behavioral boundaries is imperfect.
(a) Corpus findings on how event boundaries are computed (surprise, context change, goal shift); convergence across methods (b) Discrepancies with behavioral boundaries — model misspecification vs. genuine neural-behavioral divergence vs. individual variability (c) Can generative AI (Sora, Veo for video; LLMs for narratives) produce parametrically-controlled naturalistic stimuli for causal tests?
Synthesize with citations.
Why: HMM event boundaries는 이론적 artifact일 가능성. Generative AI로 causal manipulation 가능.
The corpus shows strong hippocampal-cortical dialogue during naturalistic encoding (Kwon 2025, Chen-Cohen 2022).
(a) Established mechanisms — event-boundary-triggered consolidation, continuous replay, schema integration, hippocampal-DMN coupling (b) Generalization across movies vs. narratives vs. real-life; mental reinstatement role (c) AI-based memory models (episodic memory in transformers, Hopfield associative memory, RAG architectures) informed by naturalistic fMRI
Synthesize with citations.
Why: 연속 경험의 이산 기억 변환 메커니즘 (event boundary → consolidation) 이 AI 메모리 아키텍처에 영감.
Answer file: ../answers/Q08_memory_encoding_and_reinstatement.md
Predictive coding is the dominant theoretical framework but naturalistic validation remains patchy.
(a) Naturalistic evidence — expectation violations, surprise encoding, top-down modulation, hierarchical prediction errors, Caucheteux/King 2023-style multi-timescale predictions (b) Bayesian surprise vs. simpler novelty/salience — free energy vs. mutual information vs. KL divergence operationalization (c) LLMs with controllable surprisal (varying entropy of predicted words) to disambiguate predictive-coding mechanisms
Synthesize with citations.
Why: Predictive coding theory는 광범위 주장. Naturalistic으로 실증적 구분 가능성 (Caucheteux 2023 layer-specific 증거).
LLM embeddings strongly predict cortical activity during language processing (Schrimpf 2021, Caucheteux 2023, Jain 2024).
(a) Scope and strength — which brain areas, which linguistic levels (phonemic/syntactic/semantic/discourse), which model families (transformer/RNN/cognitive) (b) Shallow (shared statistics) vs. deep (shared computation) alignment debate (c) Stringent tests — scaling laws, cross-family comparisons (transformer vs. SSM/Mamba vs. cognitive), lesion experiments, counterfactual prompts
Synthesize with citations.
Why: 분야의 가장 큰 논쟁. LLM-brain alignment가 우연한 통계 아닌 실제 계산 공유인지 stringent test 필요.
Toward a "Brain Foundation Model" trained on naturalistic fMRI that generalizes across subjects/stimuli/tasks.
(a) Corpus progress — large datasets (HCP 7T, UK Biobank, NSD, NNDb), spatiotemporal transformers, hyperalignment pretraining (b) Architecture/data requirements — transformer vs. GNN vs. state-space, pretraining objectives, scaling laws (c) Closest current work and gap to a true Brain FM; roadmap
Synthesize with citations.
Why: Brain foundation models은 분야 미래. 데이터/아키텍처 요구사항 체계적 매핑.
Answer file: ../answers/Q11_brain_foundation_models.md
Brain-to-text (Tang/Huth 2023), brain-to-image (MindEye, Takagi 2023), brain-to-video (Chen 2024) have matured.
(a) Current state and limits of generative decoding from naturalistic stimuli; fidelity benchmarks (b) Fundamental limits — modality-independent semantic meaning, personal/autobiographical interpretations, unconscious content, privacy implications (c) What decoding success/failure teaches about cortical representational format; generative AI + brain data joint tool
Synthesize with citations.
Why: Decoding은 brain reading 현실화. 한계/윤리 이슈 병행 검토 필수.
Answer file: ../answers/Q12_generative_decoding.md
Naturalistic neuroscience reveals brain computations (temporal hierarchies, event segmentation, predictive dynamics, hippocampal replay, attention prioritization).
(a) Corpus findings applicable to AI — biological timescale ratios, event-segmented processing, hippocampal replay-inspired memory, attention-modulated feedback (b) Robust transferable principles vs. species/tissue-specific (c) What would a "naturalistic-inspired foundation model" look like — comparison with brain-agnostic large models on efficiency, alignment, interpretability
Synthesize with citations.
Why: Neuro → AI 양방향. Event segmentation, timescale hierarchy 등을 LLM 아키텍처에 주입 가능성.
Answer file: ../answers/Q13_neurally_inspired_ai.md
The corpus shows naturalistic features consistently outperform resting-state for psychiatric/trait prediction.
(a) Candidate mechanisms — signal amplification (broader activation), ecological specificity (real-world match), better SNR (controlled state), individual-engagement variance amplification (b) Which disorders/traits benefit most (autism, ADHD, depression, psychosis, anxiety, personality); effect-size comparisons (c) AI-driven adaptive stimulus selection personalizing diagnostic content to maximize individual biomarker signal
Synthesize with citations.
Why: Naturalistic의 임상 우위는 반복 관찰 but 기전 불분명. 기전 파악이 stimulus design 가이드.
Answer file: ../answers/Q14_naturalistic_vs_restingstate_clinical.md
Longitudinal naturalistic fMRI is severely underrepresented (<2 corpus papers).
(a) Test-retest stability metrics for ISC, fingerprinting, event segmentation across sessions/weeks/years (b) Methodological challenges — habituation, stimulus repetition effects, dose-response confounds, developmental change interference (c) AI solutions — adaptive stimulus selection, brain-state-conditioned content generation, synthetic longitudinal benchmarks, digital twins
Synthesize with citations.
Why: 종단 연구 부족은 분야 전체 한계. AI가 habituation/repetition 문제 해결할 수 있는지 탐구.
Movie-watching brain fingerprints (Finn, Vanderwal, Rosenberg) identify individuals with >90% accuracy.
(a) Composition — stable trait (personality, IQ), state-dependent engagement, stimulus-specific interpretation, noise residual; which networks contribute (b) Decomposition studies — multi-stimulus, multi-session designs separating trait/state/stimulus (c) AI methods (multi-task learning, disentangled representations, causal inference) for separating factors
Synthesize with citations.
Why: Fingerprinting이 강력하지만 해석 불명. Trait/state/stimulus 분해는 disentangled representation의 brain application.
Answer file: ../answers/Q16_movie_watching_fingerprints.md
When do mature naturalistic processing hierarchies develop?
(a) Corpus findings on infant/child/adolescent naturalistic fMRI (Ellis 2025 infant visual cortex, Cohen 2022 story-evoked responses, Tripathy 2024 adults vs. children) (b) Earlier biomarkers for atypical development (ASD, language delay) than traditional tasks? (c) AI simulations of neural development (NeuroGPT, lifespan-trained brain FMs) to illuminate what matures when
Synthesize with citations.
Why: Naturalistic은 영유아 연구에 특히 강력 (비협조 피험자). 조기 biomarker 가능성 클리니컬 임팩트.
Answer file: ../answers/Q17_developmental_trajectory.md
Marmoset/macaque naturalistic fMRI has grown rapidly (14 corpus papers, 5 in 2024-2025).
(a) Cross-species homologies — temporal hierarchies, movie-evoked connectivity, social/face processing, attention networks (b) Methodological challenges — stimulus adaptation, cross-species alignment, scanner tech, training (c) Multimodal foundation models as shared referential embedding for human + non-human primate naturalistic data
Synthesize with citations.
Why: Comparative naturalistic fMRI는 침습적 실험(lesion, optogenetics)의 인간 결과 번역 통로. Foundation model이 shared embedding 제공.
The corpus stimuli are heavily Western, English-language, adult-media-focused.
(a) Corpus acknowledgment of cultural/demographic stimulus bias; cross-cultural replications (b) Consequences for neural "universals" vs. culturally-specific findings; global clinical translation (c) AI-generated naturalistic stimuli (multilingual LLM narratives, culturally-diverse Sora/Veo video, style-transfer film) — ethical considerations
Synthesize with citations.
Why: WEIRD 편향은 분야 건전성 문제. AI-generated stimuli가 해결책이지만 윤리/생태학적 타당성 주의.
Naturalistic fMRI remains scanner-constrained (fixed position, 2D display, isolated viewing).
(a) Extensions — VR-in-scanner, hyperscanning (Schippers, Naci, Ramseyer), real-time neurofeedback, ambulatory paradigms (b) Technical frontiers — 7T + VR, portable MEG, ultra-low-field MRI, wearable neuroimaging (c) AI as enabler — adaptive content, closed-loop stimulation, synthetic controls, cross-modality neural proxies (EEG-fMRI fusion)
Synthesize with citations.
Why: Scanner 한계는 naturalistic fMRI의 역설. VR/wearable 기술과 AI 결합으로 real-world 확장.
5개씩 병렬 (NotebookLM rate limit 고려):
queries_batch = [q1, q2, q3, q4, q5]
for q in queries_batch:
result = mcp__notebooklm__notebook_query(
notebook_id="d9265824-3383-4fd4-8d17-03512a338ee5",
query=q,
timeout=180
)
# save to tool-results/# 쿼리 전 설정
mcp__notebooklm__chat_configure(
notebook_id=nb_id,
response_length="longer" # 기본값의 1.5-2배 길이
)각 쿼리 응답 품질 체크:
- 답변 길이 >3 KB
- Citations ≥5
- 3-part 구조 유지 ((a)(b)(c))
- 외부 지식 명시적 구분 ("while the sources do not contain...")
새 쿼리 추가 시 → 06-extending.md 참조.
| # | English | Korean |
|---|---|---|
| 1 | Statistical Inference | 통계적 추론 |
| 2 | Stimulus Standardization | 자극 표준화 |
| 3 | Motion & Artifact | 동작·아티팩트 |
| 4 | Temporal Hierarchy | 시간 계층 |
| 5 | Shared vs Idiosyncratic | 공유 vs 개별 부호 |
| 6 | Multimodal Integration | 다중 모달 통합 |
| 7 | Event Segmentation | 사건 분할 |
| 8 | Memory Encoding | 기억 공고화 |
| 9 | Predictive Coding | 예측 부호화 |
| 10 | LLM-Brain Alignment | LLM-뇌 정합 |
| 11 | Brain Foundation Models | 뇌 기초 모델 |
| 12 | Generative Decoding | 생성적 디코딩 |
| 13 | Neurally-Inspired AI | 신경 영감 AI |
| 14 | Clinical Superiority | 임상 우위 |
| 15 | Longitudinal Tracking | 종단 추적 |
| 16 | Fingerprint Interpretation | 지문 해석 |
| 17 | Developmental Trajectory | 발달 궤적 |
| 18 | Cross-Species Homologies | 종간 정합 |
| 19 | Cultural Bias | 문화적 편향 |
| 20 | Beyond Movies | 영화 너머 |