-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathIITP_Proposal_Full.html
More file actions
217 lines (214 loc) · 20.7 KB
/
IITP_Proposal_Full.html
File metadata and controls
217 lines (214 loc) · 20.7 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<title>IITP Proposal Export</title>
<style>
body {
font-family: 'Malgun Gothic', 'Apple SD Gothic Neo', sans-serif;
line-height: 1.6;
max_width: 800px;
margin: 0 auto;
padding: 20px;
}
h1, h2, h3 { color: #2c3e50; }
code { background: #f4f4f4; padding: 2px 5px; border-radius: 3px; }
img { max_width: 100%; }
blockquote { border-left: 4px solid #aaf; padding-left: 10px; color: #555; }
</style>
</head>
<body>
<!-- FILE: 00_executive_summary.md -->
<h1>요약문: 인지적 그라운딩을 위한 DIVER-Neuro 파운데이션 모델</h1>
<h2>연구개발의 필요성 및 중요성</h2>
<p>현재 AI(LLM)는 방대한 텍스트를 처리하지만, 물리적 실체(Body)와 생물학적 생존 본능(Survival)이 결여되어 있다. 이는 진정한 의미의 **"자율성(Autonomy)"**과 <strong>"에너지 효율성(Energy Efficiency)"</strong> 달성을 가로막는 근본적인 한계이다.
본 연구는 이를 극복하기 위해 **뇌의 기억 기전(Hippocampus)**과 **신체성(Embodiment)**을 결합한 차세대 파운데이션 모델 **"DIVER-Neuro"**를 제안한다.</p>
<h2>핵심 해결 수단 (Core Technology)</h2>
<h3>A. Two-Part Model: Titans Memory & Dual Encoders</h3>
<p>최신 구글 DeepMind의 **Titans 아키텍처(Behrens et al., 2025)**를 뇌-컴퓨터 인터페이스(BCI)에 최초로 도입한다.</p>
<ol>
<li><strong>Sensory-Motor Encoder</strong>: 시각(ViT)과 신체 움직임(Proprioception)을 처리.</li>
<li><strong>Titans Memory Core</strong>: **'Surprisal(예측 오차)'**이 높은 정보만을 선택적으로 장기 기억에 저장(Test-time Training)하여, 기존 Transformer 대비 **100배 이상의 긴 문맥(Context)**을 에너지 효율적으로 처리한다.</li>
</ol>
<h3>B. Allostatic Neuro-Twin (알로스태틱 뉴로-트윈)</h3>
<p>단순한 반응형(Reactive) 항상성 유지를 넘어, 미래의 위험을 예측하고 사전에 에너지를 조절하는 <strong>"알로스태시스(Predicitive Allostasis)"</strong> 모델을 구현한다. 이는 사용자의 생체 신호(HRV, EEG)와 환경 정보를 통합하여 **"최적의 생존 에너지 경로(Neural Manifold)"**를 탐색한다.</p>
<h2>연구의 독창성 및 우수성 (Excellence)</h2>
<ul>
<li><strong>과학적 우위</strong>: 단순한 뇌파 분류를 넘어, <strong>BrainMamba(Wang et al., 2024)</strong> 등 최신 State Space Model(SSM) 기술을 적용하여 뇌 동역학의 **장기 의존성(Long-term Dependency)**을 규명함.</li>
<li><strong>기술적 우위</strong>: <strong>Tubularity(관형 구조)</strong> 이론(Bertram et al., 2026)을 기반으로, 노이즈가 심한 일상 환경에서도 강건한(Robust) 인터페이스를 제공함.</li>
</ul>
<h2>기대 효과</h2>
<p>본 연구는 "보고(Vision), 느끼고(Proprioception), 기억하는(Titans)" 완전한 신체화 AI를 통해, <strong>디지털 헬스케어(정신건강)</strong> 및 <strong>휴머노이드 로봇(적응형 제어)</strong> 분야의 기술적 난제를 획기적으로 해결할 것이다.</p>
<!-- FILE: 01_architecture.md -->
<h1>듀얼 인코더 & Titans 통합 아키텍처 (Embodied Neuro-AI Architecture)</h1>
<h2>개요: 신체화된 신경 인공지능 (Embodied Neuro-AI)</h2>
<p>본 제안서는 인간의 인지 과정이 뇌 내부 연산뿐만 아니라 신체(Body)와 환경(Environment)의 상호작용 속에서 발현된다는 <strong>"Embodied AI"</strong> 철학을 따름. 시각/청각 등 외부 감각(External Sense)과 뇌파/고유감각 등 내부 신호(Internal Signal)를 통합하여, 인간 수준의 적응성과 강건성을 갖춘 **"DIVER-Neuro 파운데이션 모델"**을 제안함.</p>
<p><img src="../../05_figures/dual_encoder_titans_architecture.png" alt="Dual Encoder Titans Architecture" /></p>
<h2>듀얼 인코더 시스템 (Dual Encoders)</h2>
<p>서로 다른 물리적 특성을 가진 두 가지 데이터 스트림 처리를 위해 특화된 <strong>Dual Encoder</strong> 구조를 채택함.</p>
<h3>Sensory-Motor & Interoceptive Encoder (OmniField)</h3>
<p><strong>OmniField</strong> [@valencia2025omnifield]는 시각/청각뿐만 아니라 <strong>내수용 감각(Interoception)</strong> 신호를 통합하는 4D Neural Field입니다.</p>
<ul>
<li><strong>External</strong>: Camera(Ego-centric), Mic(Audio)의 비정형 데이터를 처리함.</li>
<li><strong>Internal (Allostasis)</strong>: 위장 팽창(Gastric Distension)과 같은 생체 신호를 연속적인 **내부 상태 변수(Internal State Variable)**로 인코딩합니다.
<ul>
<li><strong>Biological Grounding</strong>: 김성연 교수팀[@kim2020neural]의 연구에 기반하여, <strong>NTS $\rightarrow$ PVH $\rightarrow$ aIC</strong> 신경 회로를 모사합니다. 위장 팽창 신호를 단순 자극이 아닌 '적분형(Slow Integrator)' 정보로 처리하여 행동 모드(Exploration vs. Exploitation) 전환의 기준점으로 삼습니다.</li>
</ul>
</li>
<li><strong>Latent Sampler</strong>: 연속적인 OmniField의 출력을 Titans가 처리 가능한 토큰($z_t$)으로 변환하기 위해, <strong>좌표 기반 샘플링(Coordinate Sampling)</strong> 층을 둠. 이 과정에서 Han et al.[@han2024simulated]의 <strong>Simulated Annealing</strong> 기법을 초기화 단계에 적용하여, 최적의 특징점(Feature Points)을 포착하고 Local Minima를 방지함.</li>
</ul>
<h3>Brain Spatiotemporal Encoder (뇌 시공간 인코더)</h3>
<p>뇌 활동(EEG/iEEG)의 복잡한 시공간적 동역학(Spatiotemporal Dynamics)을 모델링함.</p>
<ul>
<li><strong>Liquid Time-Constant (LTC) Networks</strong>: 불규칙하고 연속적인 뇌 신호 처리를 위해, 입력에 따라 유동적으로 변하는 <strong>Time-constant</strong>를 가진 미분방정식 기반 **LTC (Neural ODE)**를 도입함[2].</li>
<li><strong>장점</strong>: 노이즈가 많은 생체 신호 처리 시 기존 RNN/LSTM 대비 월등한 강건성을 보이며, 인과관계(Causality) 추론에 유리함.</li>
</ul>
<h2>Titans Memory: 통합과 선택적 주의 (The Integrator)</h2>
<p>이종(Heterogeneous) 표상은 <strong>Titans Memory Core</strong>에서 통합됨.</p>
<ul>
<li><strong>Neural Memory Module (MIRAS Framework)</strong>: Google의 최신 <strong>MIRAS (Memory as Context)</strong> 프레임워크[@behrens2025titans]에 기반하여, 단순 저장이 아닌 **"Surprise Metric (놀라움 척도)"**과 **"Momentum (관성)"**을 이용해 능동적으로 중요 기억을 선별함.</li>
<li><strong>Global Neural Workspace (GNW)</strong>: Dehaene 등[5]의 GNW 이론에 따라, Titans는 감각/뇌 정보가 경쟁·통합되는 "의식적 작업 공간" 역할을 수행함.</li>
<li><strong>Graph-Informed Reasoning</strong>: Moontae Lee(LG AI Research) 심층 추론 방법론[@lee2024strategic]을 적용, Titans 메모리에 저장된 에피소드들을 그래프 구조로 연결하여 단순 회상(Recall)을 넘어선 **전략적 추론(Strategic Reasoning)**을 수행함.</li>
<li><strong>Tubular Manifold Alignment</strong>: 통합 과정에서 모델은 Bertram 등[4]의 **"Tubularity(관형 구조)"**를 유지하도록 학습되어 강건성을 확보함.</li>
</ul>
<h2>결론 및 차별점</h2>
<p>본 아키텍처는 단순 멀티모달 결합(Early Fusion)을 넘어, <strong>LTC 기반 생물학적 동역학 모델링</strong>과 <strong>Titans 기반 능동적 메모리 관리</strong>를 결합한 세계 최초 시도임. 이를 통해 IITP RFP가 요구하는 "인간 수준 인지 능력을 갖춘 뇌 내재화 모델"을 실현함.</p>
<!-- FILE: 02_methodology.md -->
<h1>제2장: 검증 및 방법론 (Validation & Methodology)</h1>
<h2>2.1 Tubularity & Manifold Alignment (검증 지표)</h2>
<p>본 연구는 AI 모델이 단순한 패턴 매칭을 넘어, 인간 뇌와 유사한 **"Neural Manifold(신경 다양체)"**를 형성하는지 검증하기 위해 Bertram et al.[4]이 제안한 혁신적 지표인 **"Tubularity"**를 도입함.</p>
<h3>2.1.1 Neural Manifold Alignment</h3>
<p>이종(Heterogeneous) 데이터(시각, 언어, 뇌파)가 통합될 때, 각 모달리티의 잠재 공간(Latent Space)이 서로 엉키지 않고 매끄럽게 정렬되어야 강건한 추론이 가능함.</p>
<ul>
<li><strong>Tubularity Metric</strong>: 학습된 궤적(Trajectory)이 국소적으로 원통형(Tubular) 구조를 유지하는지 측정함. 이는 모델이 노이즈에 얼마나 강건한지(Robustness)를 나타내는 핵심 지표임.</li>
<li><strong>검증 목표</strong>: DIVER-Neuro 모델의 Tubularity Score가 기존 Transformer 기반 모델 대비 30 percent 이상 향상됨을 입증함.</li>
</ul>
<h1>실험 및 검증 방법론 (Methodology)</h1>
<h2>데이터셋 구축 (Dataset Construction)</h2>
<p>본 연구는 뇌 활동(Brain)과 외부 환경(Vision)을 동시에 포함하는 멀티모달 데이터셋을 활용함.</p>
<h3>Toy Dataset: Moving MNIST + Synthetic EEG</h3>
<ul>
<li><strong>Visual</strong>: $28 \times 28$ Moving MNIST (2 digits).</li>
<li><strong>Brain (Synthetic)</strong>: 비선형 Izhikevich 뉴런 모델을 사용하여, Visual Stimulus의 속도와 위치에 반응하는 가상의 EEG 생성.</li>
<li><strong>목적</strong>: Dual Encoder의 기본 동작 및 Latent Alignment 검증.</li>
</ul>
<h3>Main Dataset: BCM-V (Brain-Computer Interface Multimodal Video)</h3>
<ul>
<li><strong>구성</strong>: 피험자가 1인칭 시점(Ego-centric) 영상을 시청할 때 측정된 128채널 EEG 및 Eye-tracking 데이터.</li>
<li><strong>특징</strong>: 자연스러운 환경에서의 시각-뇌 활동 상관관계를 포함.</li>
</ul>
<h2>학습 전략 (Training Strategy)</h2>
<h3>Phase 1: Pre-training (Brain-Visual Alignment)</h3>
<ul>
<li><strong>Contrastive Learning (CLIP Style)</strong>: OmniField(Visual)와 Brain Encoder(LTC)의 Latent Vector $z_v, z_b$ 간의 코사인 유사도를 최대화함.</li>
<li><strong>Masked Modeling (MAE Style)</strong>: Brain Encoder의 일부 타임스텝을 마스킹하고 복원하도록 학습하여 시간적 문맥(Temporal Context) 학습.</li>
</ul>
<h3>Phase 2: Titans Memory Integration</h3>
<ul>
<li><strong>Surprise-based Update</strong>: 정렬된 Latent Vector를 Titans에 입력. 예측 오차(Prediction Error)가 높은 "놀라운(Surprising)" 정보만 Long-term Memory에 저장.</li>
<li><strong>Free Energy Minimization</strong>: MIRAS의 <strong>Surprise Metric</strong>은 수학적으로 Friston의 **Free Energy(예측 불확실성)**와 동등함[@friston2010free]. Titans는 이 Free Energy를 최소화하는 방향으로 메모리를 갱신함으로써, 생물학적 항상성(Homeostasis) 원리를 구현함.</li>
</ul>
<h3>Brain-Tuning: Semantic Alignment</h3>
<p>DIVER의 fMRI 데이터를 사용하여 Titans의 상위 레벨 표상을 의미론적으로 정렬합니다 [@benara2025braintuning].</p>
<h3>Biological Validation: Gastric Interoception (Murine Data)</h3>
<p>Titans의 기억 메커니즘이 실제 생물학적 회로와 유사하게 작동하는지 검증하기 위해, <strong>김성연 교수팀의 동물 실험 데이터</strong>를 활용합니다 [@kim2020neural].</p>
<ul>
<li><strong>Data Source</strong>: 생쥐(Mouse)의 <strong>PB Pdyn 뉴런 Two-photon Calcium Imaging</strong> 데이터 (위장 팽창 시의 단일 뉴런 활성).</li>
<li><strong>Hypothesis</strong>: Titans의 <strong>Surprise Metric</strong>은 생쥐의 PB Pdyn 뉴런이 보여주는 <strong>Sustained Activity (Integrator)</strong> 패턴과 높은 상관관계를 보여야 합니다.</li>
<li><strong>Validation</strong>: 위장 팽창(Gastric Distension) 시뮬레이션 환경에서, Titans의 Memory State 변화가 실제 신경 데이터의 역동성과 일치하는지 정량적으로 분석합니다. 이는 AI가 단순한 정보 저장이 아닌, 생체와 유사한 **'내부 상태 적분(Internal State Integration)'**을 수행함을 증명합니다.</li>
</ul>
<h2>실험 및 검증 계획 (Validation Plan)</h2>
<ul>
<li><strong>Phase 1</strong>: Toy Dataset(Moving MNIST + Synthetic EEG)을 이용한 개념 증명.</li>
<li><strong>Phase 2</strong>: BCM-V(뇌-컴퓨터 인터페이스) 데이터셋을 활용한 대규모 학습 및 Tubularity 측정.</li>
</ul>
<!-- FILE: 03_allostasis.md -->
<h1>내수용 감각장 및 알로스태틱 뉴로트윈 (Allostatic Neuro-Twin)</h1>
<h2>개요: 예측적 생존 본능 (Predictive Allostasis)</h2>
<p>기존의 AI 에이전트나 웰니스 앱은 문제가 발생한 후 반응하는 <strong>Homeostasis(항상성, Reactive)</strong> 모델에 그쳤음. 본 과제는 시스템이 변화하는 미래 환경을 예측하고 사전에 에너지를 조절하는 <strong>Allostasis(알로스태시스, Predictive Regulation)</strong> 개념을 도입하여, 에이전트의 생존성을 극대화함.</p>
<p><img src="../../05_figures/allostasis_vs_homeostasis_diagram.png" alt="Interoceptive Allostasis vs Homeostasis" /></p>
<h2>시스템 구조: Neuro-Twin Loop</h2>
<p>사용자의 신체 및 환경 데이터를 실시간으로 미러링하는 **"Digital Neuro-Twin"**을 구축함.</p>
<h3>Ubiquitous & Visceral Sensing (입력)</h3>
<ul>
<li><strong>Visceral Interoception</strong>: 김성연 교수팀[@kim2020neural]의 연구에 기반하여, 위장 팽창(Gastric Distension)과 같은 내부 장기 신호를 <strong>NTS $\rightarrow$ PVH $\rightarrow$ aIC</strong> 경로를 통해 실시간으로 추적함. 이는 단순한 포만감을 넘어, 에이전트의 **기초 에너지 수준(Basal Energy Level)**을 정의하는 핵심 변수로 작용함.</li>
<li><strong>Passive Sensing</strong>: 스마트워치/글래스 등 웨어러블 기기를 통해 심박 변이도(HRV), 피부 전도도(EDA) 등 자율신경계 신호를 수동적으로 수집함.</li>
<li><strong>Neural Proxy based on Gastric State</strong>: 위장 상태(Gastric State)가 뇌의 각성(Arousal) 및 행동 모드(Exploration vs. Exploitation)를 조절한다는 이론[@berntson2021neural]을 적용하여, 내장 신호로부터 뇌의 거시적 상태를 역추론함.</li>
</ul>
<h3>Energy Landscape Inference (처리)</h3>
<p>뇌 상태를 고차원 <strong>"Energy Landscape(에너지 지형)"</strong> 상의 좌표로 매핑함.</p>
<ul>
<li><strong>Pathological Attractor</strong>: 우울, 불안, 또는 시스템 과부하 상태를 '빠져나오기 힘든 깊은 골짜기(Attractor)'로 정의함.</li>
<li><strong>Allostatic Load</strong>: 현재 상태가 지속될 경우 시스템에 가해질 누적 부하(Load)를 예측 시뮬레이션함.</li>
</ul>
<h3>Predictive Nudge (개입)</h3>
<p>시스템이나 사용자가 병리적 끌개(Attractor)로 진입하기 전, 선제적으로 개입함.</p>
<ul>
<li><strong>Intervention</strong>: 조명, 온도 조절 또는 사용자에게 알림(Nudge)을 제공하여, 에너지 지형 상의 '안전 지대'로 궤적을 수정함[Psychology Today 2025].</li>
</ul>
<h2>기대 효과</h2>
<p>단순한 헬스케어를 넘어, 극한 환경(재난, 우주 등)에서 인간과 AI 시스템의 공생(Symbiosis)을 가능케 하는 **"Energy-Efficient Robust Intelligence"**를 구현함.</p>
<!-- FILE: 04_validation.md -->
<h1>이론적 타당성 및 성능 목표 (Validation)</h1>
<h2>Theoretical Superiority (이론적 우수성)</h2>
<p>본 제안서는 단순한 실험적 시도가 아닌, 2024-2025년 발표된 최신 <strong>SOTA(State-of-the-Art)</strong> 연구 결과들에 기반하여 설계되었다.</p>
<h3>Transformer vs. Titans/SSM in Brain</h3>
<p>기존 Transformer(ViT)는 $O(N^2)$의 연산 복잡도로 인해 긴 시계열의 생체 신호를 처리하는 데 한계가 있다.</p>
<ul>
<li><strong>근거</strong>: <strong>BrainMamba (Wang et al., 2024)</strong> 연구에 따르면, SSM 기반 모델이 Transformer 대비 <strong>메모리 사용량 50 percent 절감</strong>, <strong>추론 속도 5배 향상</strong>을 달성하면서도 EEG 분류 정확도는 더 높았다.</li>
<li><strong>전략</strong>: 본 연구는 BrainMamba의 SSM 코어를 **Titans Memory (Behrens et al., 2025)**로 확장하여, 단순 분류를 넘어 **"생애 전주기 기억(Lifelong Learning)"**이 가능한 아키텍처를 구현한다.</li>
</ul>
<h2>목표 성능 지표 (Performance Metrics)</h2>
<h3>1. Robustness: Tubularity Score</h3>
<ul>
<li><strong>정의</strong>: Neural Manifold가 노이즈(외란) 속에서도 고유의 위상학적 구조(Topology)를 유지하는 정도.</li>
<li><strong>Target</strong>: Baseline (ResNet/ViT) 대비 <strong>+30 percent 향상</strong>. (근거: Bertram et al., 2026).</li>
</ul>
<h3>2. Efficiency: Effective Context Length</h3>
<ul>
<li><strong>정의</strong>: 손실 없이 역전파 가능한 최대 시퀀스 길이.</li>
<li><strong>Target</strong>: <strong>1M Tokens</strong> (Titans Memory 적용 시). 기존 RNN(1k), Transformer(8k-32k)를 압도함.</li>
</ul>
<h3>3. Generalization: Zero-shot Adaptation</h3>
<ul>
<li><strong>정의</strong>: 학습하지 않은 새로운 피험자(New Subject)에 대한 적응 능력.</li>
<li><strong>Target</strong>: 학습 데이터 없이도 <strong>Test-time Training</strong>을 통해 즉시 80 percent 이상의 정확도 확보.</li>
</ul>
<h2>결론</h2>
<p>DIVER-Neuro 모델은 <strong>Titans의 기억 능력</strong>과 <strong>SSM의 효율성</strong>을 **신체화(Embodiment)**된 에이전트에 통합함으로써, 차세대 AI의 새로운 표준(Standard)을 제시한다.</p>
<!-- FILE: 99_red_team_review.md -->
<h1>🛡️ Red Team Review Report (Expert Persona)</h1>
<p><strong>Reviewer</strong>: Prof. Neuro-AI
<strong>Date</strong>: 2026-02-03
<strong>Target</strong>: DIVER-Neuro Architecture & Methodology (Merged Drafts)</p>
<h2>Overall Impression</h2>
<p>The proposal incorporates cutting-edge components (<strong>Titans/MIRAS</strong>, <strong>OmniField</strong>, <strong>Brain Tuning</strong>) and demonstrates high ambition ("Excellence"). However, the rapid integration of these SOTA models creates "Frankenstein Risks" where components are conceptually stacked but mechanistically disconnected.</p>
<h2>Coherence Gaps (Critical)</h2>
<h3>🚨 Gap 1: The "Field-to-Token" Disconnect</h3>
<ul>
<li><strong>Structure</strong>: The Sensory Encoder uses <strong>OmniField</strong> (Continuous Neural Field). The Memory Core uses <strong>Titans</strong> (Discrete/Token-based Sequence Model).</li>
<li><strong>Critique</strong>: How does the continuous field output of OmniField become a discrete token for Titans? If this "Tokenizer" or "Sampler" step is missing, the architecture is broken.</li>
<li><strong>Action</strong>: Explicitly define a <strong>"Coordinate Sampler"</strong> or <strong>"Latent Tokenizer"</strong> layer between OmniField and Titans.</li>
</ul>
<h3>🚨 Gap 2: Brain Tuning Data Feasibility</h3>
<ul>
<li><strong>Structure</strong>: Methodology 2.3 proposes <strong>Brain-Tuning</strong> using Toneva's method (Language-to-Brain alignment).</li>
<li><strong>Critique</strong>: Toneva's work relies on Text-fMRI pairs. DIVER deals with "Sensory-Motor" data. Do you have "Motor-fMRI" pairs? Or are you tuning only the "Concept/Language" part of Titans?</li>
<li><strong>Action</strong>: Clarify that Brain Tuning applies primarily to the <strong>"Semantic/Concept Latent Space"</strong> of Titans, leveraging the Language-Brain datasets (DIVER-Lang subset), while Sensory parts use standard reconstruction loss.</li>
</ul>
<h3>🚨 Gap 3: Missing "Simulated Annealing" Strategy</h3>
<ul>
<li><strong>Structure</strong>: Han et al. (2024) "Simulated Annealing in Early Layers" is cited.</li>
<li><strong>Critique</strong>: It appears in the bibliography but is <strong>nowhere in the text</strong>. Why cite it if you don't use it?</li>
<li><strong>Action</strong>: Integrate it into <strong>Phase 1 Optimization Strategy</strong> or removing it if irrelevant. (Suggestion: Use it for the <strong>OmniField Initialization</strong> to prevent local minima).</li>
</ul>
<h3>🚨 Gap 4: MIRAS vs. Allostasis</h3>
<ul>
<li><strong>Structure</strong>: MIRAS uses "Surprise Metric". Allostasis uses "Free Energy Minimization".</li>
<li><strong>Critique</strong>: These are mathematically similar but terminology is split.</li>
<li><strong>Action</strong>: Explicitly state: <strong>"MIRAS Surprise $\approx$ Free Energy (Prediction Error)."</strong> Unify the narrative.</li>
</ul>
<h2>Conclusion</h2>
<p>The "Excellence" is high, but "Rigour" is slightly loose. Address Gap 1 (Sampler) and Gap 4 (Unification) immediately. Gap 2 & 3 are secondary but strengthen the proposal.</p>
</body>
</html>