-
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) ยท 20.7 KB
/
IITP_Proposal_Full.html
File metadata and controls
217 lines (214 loc) ยท 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>์์ฝ๋ฌธ: ์ธ์ง์ ๊ทธ๋ผ์ด๋ฉ์ ์ํ DIVER-Neuro ํ์ด๋ฐ์ด์
๋ชจ๋ธ</h1>
<h2>์ฐ๊ตฌ๊ฐ๋ฐ์ ํ์์ฑ ๋ฐ ์ค์์ฑ</h2>
<p>ํ์ฌ AI(LLM)๋ ๋ฐฉ๋ํ ํ
์คํธ๋ฅผ ์ฒ๋ฆฌํ์ง๋ง, ๋ฌผ๋ฆฌ์ ์ค์ฒด(Body)์ ์๋ฌผํ์ ์์กด ๋ณธ๋ฅ(Survival)์ด ๊ฒฐ์ฌ๋์ด ์๋ค. ์ด๋ ์ง์ ํ ์๋ฏธ์ **"์์จ์ฑ(Autonomy)"**๊ณผ <strong>"์๋์ง ํจ์จ์ฑ(Energy Efficiency)"</strong> ๋ฌ์ฑ์ ๊ฐ๋ก๋ง๋ ๊ทผ๋ณธ์ ์ธ ํ๊ณ์ด๋ค.
๋ณธ ์ฐ๊ตฌ๋ ์ด๋ฅผ ๊ทน๋ณตํ๊ธฐ ์ํด **๋์ ๊ธฐ์ต ๊ธฐ์ (Hippocampus)**๊ณผ **์ ์ฒด์ฑ(Embodiment)**์ ๊ฒฐํฉํ ์ฐจ์ธ๋ ํ์ด๋ฐ์ด์
๋ชจ๋ธ **"DIVER-Neuro"**๋ฅผ ์ ์ํ๋ค.</p>
<h2>ํต์ฌ ํด๊ฒฐ ์๋จ (Core Technology)</h2>
<h3>A. Two-Part Model: Titans Memory & Dual Encoders</h3>
<p>์ต์ ๊ตฌ๊ธ DeepMind์ **Titans ์ํคํ
์ฒ(Behrens et al., 2025)**๋ฅผ ๋-์ปดํจํฐ ์ธํฐํ์ด์ค(BCI)์ ์ต์ด๋ก ๋์
ํ๋ค.</p>
<ol>
<li><strong>Sensory-Motor Encoder</strong>: ์๊ฐ(ViT)๊ณผ ์ ์ฒด ์์ง์(Proprioception)์ ์ฒ๋ฆฌ.</li>
<li><strong>Titans Memory Core</strong>: **'Surprisal(์์ธก ์ค์ฐจ)'**์ด ๋์ ์ ๋ณด๋ง์ ์ ํ์ ์ผ๋ก ์ฅ๊ธฐ ๊ธฐ์ต์ ์ ์ฅ(Test-time Training)ํ์ฌ, ๊ธฐ์กด Transformer ๋๋น **100๋ฐฐ ์ด์์ ๊ธด ๋ฌธ๋งฅ(Context)**์ ์๋์ง ํจ์จ์ ์ผ๋ก ์ฒ๋ฆฌํ๋ค.</li>
</ol>
<h3>B. Allostatic Neuro-Twin (์๋ก์คํํฑ ๋ด๋ก-ํธ์)</h3>
<p>๋จ์ํ ๋ฐ์ํ(Reactive) ํญ์์ฑ ์ ์ง๋ฅผ ๋์ด, ๋ฏธ๋์ ์ํ์ ์์ธกํ๊ณ ์ฌ์ ์ ์๋์ง๋ฅผ ์กฐ์ ํ๋ <strong>"์๋ก์คํ์์ค(Predicitive Allostasis)"</strong> ๋ชจ๋ธ์ ๊ตฌํํ๋ค. ์ด๋ ์ฌ์ฉ์์ ์์ฒด ์ ํธ(HRV, EEG)์ ํ๊ฒฝ ์ ๋ณด๋ฅผ ํตํฉํ์ฌ **"์ต์ ์ ์์กด ์๋์ง ๊ฒฝ๋ก(Neural Manifold)"**๋ฅผ ํ์ํ๋ค.</p>
<h2>์ฐ๊ตฌ์ ๋
์ฐฝ์ฑ ๋ฐ ์ฐ์์ฑ (Excellence)</h2>
<ul>
<li><strong>๊ณผํ์ ์ฐ์</strong>: ๋จ์ํ ๋ํ ๋ถ๋ฅ๋ฅผ ๋์ด, <strong>BrainMamba(Wang et al., 2024)</strong> ๋ฑ ์ต์ State Space Model(SSM) ๊ธฐ์ ์ ์ ์ฉํ์ฌ ๋ ๋์ญํ์ **์ฅ๊ธฐ ์์กด์ฑ(Long-term Dependency)**์ ๊ท๋ช
ํจ.</li>
<li><strong>๊ธฐ์ ์ ์ฐ์</strong>: <strong>Tubularity(๊ดํ ๊ตฌ์กฐ)</strong> ์ด๋ก (Bertram et al., 2026)์ ๊ธฐ๋ฐ์ผ๋ก, ๋
ธ์ด์ฆ๊ฐ ์ฌํ ์ผ์ ํ๊ฒฝ์์๋ ๊ฐ๊ฑดํ(Robust) ์ธํฐํ์ด์ค๋ฅผ ์ ๊ณตํจ.</li>
</ul>
<h2>๊ธฐ๋ ํจ๊ณผ</h2>
<p>๋ณธ ์ฐ๊ตฌ๋ "๋ณด๊ณ (Vision), ๋๋ผ๊ณ (Proprioception), ๊ธฐ์ตํ๋(Titans)" ์์ ํ ์ ์ฒดํ AI๋ฅผ ํตํด, <strong>๋์งํธ ํฌ์ค์ผ์ด(์ ์ ๊ฑด๊ฐ)</strong> ๋ฐ <strong>ํด๋จธ๋
ธ์ด๋ ๋ก๋ด(์ ์ํ ์ ์ด)</strong> ๋ถ์ผ์ ๊ธฐ์ ์ ๋์ ๋ฅผ ํ๊ธฐ์ ์ผ๋ก ํด๊ฒฐํ ๊ฒ์ด๋ค.</p>
<!-- FILE: 01_architecture.md -->
<h1>๋์ผ ์ธ์ฝ๋ & Titans ํตํฉ ์ํคํ
์ฒ (Embodied Neuro-AI Architecture)</h1>
<h2>๊ฐ์: ์ ์ฒดํ๋ ์ ๊ฒฝ ์ธ๊ณต์ง๋ฅ (Embodied Neuro-AI)</h2>
<p>๋ณธ ์ ์์๋ ์ธ๊ฐ์ ์ธ์ง ๊ณผ์ ์ด ๋ ๋ด๋ถ ์ฐ์ฐ๋ฟ๋ง ์๋๋ผ ์ ์ฒด(Body)์ ํ๊ฒฝ(Environment)์ ์ํธ์์ฉ ์์์ ๋ฐํ๋๋ค๋ <strong>"Embodied AI"</strong> ์ฒ ํ์ ๋ฐ๋ฆ. ์๊ฐ/์ฒญ๊ฐ ๋ฑ ์ธ๋ถ ๊ฐ๊ฐ(External Sense)๊ณผ ๋ํ/๊ณ ์ ๊ฐ๊ฐ ๋ฑ ๋ด๋ถ ์ ํธ(Internal Signal)๋ฅผ ํตํฉํ์ฌ, ์ธ๊ฐ ์์ค์ ์ ์์ฑ๊ณผ ๊ฐ๊ฑด์ฑ์ ๊ฐ์ถ **"DIVER-Neuro ํ์ด๋ฐ์ด์
๋ชจ๋ธ"**์ ์ ์ํจ.</p>
<p><img src="../../05_figures/dual_encoder_titans_architecture.png" alt="Dual Encoder Titans Architecture" /></p>
<h2>๋์ผ ์ธ์ฝ๋ ์์คํ
(Dual Encoders)</h2>
<p>์๋ก ๋ค๋ฅธ ๋ฌผ๋ฆฌ์ ํน์ฑ์ ๊ฐ์ง ๋ ๊ฐ์ง ๋ฐ์ดํฐ ์คํธ๋ฆผ ์ฒ๋ฆฌ๋ฅผ ์ํด ํนํ๋ <strong>Dual Encoder</strong> ๊ตฌ์กฐ๋ฅผ ์ฑํํจ.</p>
<h3>Sensory-Motor & Interoceptive Encoder (OmniField)</h3>
<p><strong>OmniField</strong> [@valencia2025omnifield]๋ ์๊ฐ/์ฒญ๊ฐ๋ฟ๋ง ์๋๋ผ <strong>๋ด์์ฉ ๊ฐ๊ฐ(Interoception)</strong> ์ ํธ๋ฅผ ํตํฉํ๋ 4D Neural Field์
๋๋ค.</p>
<ul>
<li><strong>External</strong>: Camera(Ego-centric), Mic(Audio)์ ๋น์ ํ ๋ฐ์ดํฐ๋ฅผ ์ฒ๋ฆฌํจ.</li>
<li><strong>Internal (Allostasis)</strong>: ์์ฅ ํฝ์ฐฝ(Gastric Distension)๊ณผ ๊ฐ์ ์์ฒด ์ ํธ๋ฅผ ์ฐ์์ ์ธ **๋ด๋ถ ์ํ ๋ณ์(Internal State Variable)**๋ก ์ธ์ฝ๋ฉํฉ๋๋ค.
<ul>
<li><strong>Biological Grounding</strong>: ๊น์ฑ์ฐ ๊ต์ํ[@kim2020neural]์ ์ฐ๊ตฌ์ ๊ธฐ๋ฐํ์ฌ, <strong>NTS $\rightarrow$ PVH $\rightarrow$ aIC</strong> ์ ๊ฒฝ ํ๋ก๋ฅผ ๋ชจ์ฌํฉ๋๋ค. ์์ฅ ํฝ์ฐฝ ์ ํธ๋ฅผ ๋จ์ ์๊ทน์ด ์๋ '์ ๋ถํ(Slow Integrator)' ์ ๋ณด๋ก ์ฒ๋ฆฌํ์ฌ ํ๋ ๋ชจ๋(Exploration vs. Exploitation) ์ ํ์ ๊ธฐ์ค์ ์ผ๋ก ์ผ์ต๋๋ค.</li>
</ul>
</li>
<li><strong>Latent Sampler</strong>: ์ฐ์์ ์ธ OmniField์ ์ถ๋ ฅ์ Titans๊ฐ ์ฒ๋ฆฌ ๊ฐ๋ฅํ ํ ํฐ($z_t$)์ผ๋ก ๋ณํํ๊ธฐ ์ํด, <strong>์ขํ ๊ธฐ๋ฐ ์ํ๋ง(Coordinate Sampling)</strong> ์ธต์ ๋ . ์ด ๊ณผ์ ์์ Han et al.[@han2024simulated]์ <strong>Simulated Annealing</strong> ๊ธฐ๋ฒ์ ์ด๊ธฐํ ๋จ๊ณ์ ์ ์ฉํ์ฌ, ์ต์ ์ ํน์ง์ (Feature Points)์ ํฌ์ฐฉํ๊ณ Local Minima๋ฅผ ๋ฐฉ์งํจ.</li>
</ul>
<h3>Brain Spatiotemporal Encoder (๋ ์๊ณต๊ฐ ์ธ์ฝ๋)</h3>
<p>๋ ํ๋(EEG/iEEG)์ ๋ณต์กํ ์๊ณต๊ฐ์ ๋์ญํ(Spatiotemporal Dynamics)์ ๋ชจ๋ธ๋งํจ.</p>
<ul>
<li><strong>Liquid Time-Constant (LTC) Networks</strong>: ๋ถ๊ท์นํ๊ณ ์ฐ์์ ์ธ ๋ ์ ํธ ์ฒ๋ฆฌ๋ฅผ ์ํด, ์
๋ ฅ์ ๋ฐ๋ผ ์ ๋์ ์ผ๋ก ๋ณํ๋ <strong>Time-constant</strong>๋ฅผ ๊ฐ์ง ๋ฏธ๋ถ๋ฐฉ์ ์ ๊ธฐ๋ฐ **LTC (Neural ODE)**๋ฅผ ๋์
ํจ[2].</li>
<li><strong>์ฅ์ </strong>: ๋
ธ์ด์ฆ๊ฐ ๋ง์ ์์ฒด ์ ํธ ์ฒ๋ฆฌ ์ ๊ธฐ์กด RNN/LSTM ๋๋น ์๋ฑํ ๊ฐ๊ฑด์ฑ์ ๋ณด์ด๋ฉฐ, ์ธ๊ณผ๊ด๊ณ(Causality) ์ถ๋ก ์ ์ ๋ฆฌํจ.</li>
</ul>
<h2>Titans Memory: ํตํฉ๊ณผ ์ ํ์ ์ฃผ์ (The Integrator)</h2>
<p>์ด์ข
(Heterogeneous) ํ์์ <strong>Titans Memory Core</strong>์์ ํตํฉ๋จ.</p>
<ul>
<li><strong>Neural Memory Module (MIRAS Framework)</strong>: Google์ ์ต์ <strong>MIRAS (Memory as Context)</strong> ํ๋ ์์ํฌ[@behrens2025titans]์ ๊ธฐ๋ฐํ์ฌ, ๋จ์ ์ ์ฅ์ด ์๋ **"Surprise Metric (๋๋ผ์ ์ฒ๋)"**๊ณผ **"Momentum (๊ด์ฑ)"**์ ์ด์ฉํด ๋ฅ๋์ ์ผ๋ก ์ค์ ๊ธฐ์ต์ ์ ๋ณํจ.</li>
<li><strong>Global Neural Workspace (GNW)</strong>: Dehaene ๋ฑ[5]์ GNW ์ด๋ก ์ ๋ฐ๋ผ, Titans๋ ๊ฐ๊ฐ/๋ ์ ๋ณด๊ฐ ๊ฒฝ์ยทํตํฉ๋๋ "์์์ ์์
๊ณต๊ฐ" ์ญํ ์ ์ํํจ.</li>
<li><strong>Graph-Informed Reasoning</strong>: Moontae Lee(LG AI Research) ์ฌ์ธต ์ถ๋ก ๋ฐฉ๋ฒ๋ก [@lee2024strategic]์ ์ ์ฉ, Titans ๋ฉ๋ชจ๋ฆฌ์ ์ ์ฅ๋ ์ํผ์๋๋ค์ ๊ทธ๋ํ ๊ตฌ์กฐ๋ก ์ฐ๊ฒฐํ์ฌ ๋จ์ ํ์(Recall)์ ๋์ด์ **์ ๋ต์ ์ถ๋ก (Strategic Reasoning)**์ ์ํํจ.</li>
<li><strong>Tubular Manifold Alignment</strong>: ํตํฉ ๊ณผ์ ์์ ๋ชจ๋ธ์ Bertram ๋ฑ[4]์ **"Tubularity(๊ดํ ๊ตฌ์กฐ)"**๋ฅผ ์ ์งํ๋๋ก ํ์ต๋์ด ๊ฐ๊ฑด์ฑ์ ํ๋ณดํจ.</li>
</ul>
<h2>๊ฒฐ๋ก ๋ฐ ์ฐจ๋ณ์ </h2>
<p>๋ณธ ์ํคํ
์ฒ๋ ๋จ์ ๋ฉํฐ๋ชจ๋ฌ ๊ฒฐํฉ(Early Fusion)์ ๋์ด, <strong>LTC ๊ธฐ๋ฐ ์๋ฌผํ์ ๋์ญํ ๋ชจ๋ธ๋ง</strong>๊ณผ <strong>Titans ๊ธฐ๋ฐ ๋ฅ๋์ ๋ฉ๋ชจ๋ฆฌ ๊ด๋ฆฌ</strong>๋ฅผ ๊ฒฐํฉํ ์ธ๊ณ ์ต์ด ์๋์. ์ด๋ฅผ ํตํด IITP RFP๊ฐ ์๊ตฌํ๋ "์ธ๊ฐ ์์ค ์ธ์ง ๋ฅ๋ ฅ์ ๊ฐ์ถ ๋ ๋ด์ฌํ ๋ชจ๋ธ"์ ์คํํจ.</p>
<!-- FILE: 02_methodology.md -->
<h1>์ 2์ฅ: ๊ฒ์ฆ ๋ฐ ๋ฐฉ๋ฒ๋ก (Validation & Methodology)</h1>
<h2>2.1 Tubularity & Manifold Alignment (๊ฒ์ฆ ์งํ)</h2>
<p>๋ณธ ์ฐ๊ตฌ๋ AI ๋ชจ๋ธ์ด ๋จ์ํ ํจํด ๋งค์นญ์ ๋์ด, ์ธ๊ฐ ๋์ ์ ์ฌํ **"Neural Manifold(์ ๊ฒฝ ๋ค์์ฒด)"**๋ฅผ ํ์ฑํ๋์ง ๊ฒ์ฆํ๊ธฐ ์ํด Bertram et al.[4]์ด ์ ์ํ ํ์ ์ ์งํ์ธ **"Tubularity"**๋ฅผ ๋์
ํจ.</p>
<h3>2.1.1 Neural Manifold Alignment</h3>
<p>์ด์ข
(Heterogeneous) ๋ฐ์ดํฐ(์๊ฐ, ์ธ์ด, ๋ํ)๊ฐ ํตํฉ๋ ๋, ๊ฐ ๋ชจ๋ฌ๋ฆฌํฐ์ ์ ์ฌ ๊ณต๊ฐ(Latent Space)์ด ์๋ก ์ํค์ง ์๊ณ ๋งค๋๋ฝ๊ฒ ์ ๋ ฌ๋์ด์ผ ๊ฐ๊ฑดํ ์ถ๋ก ์ด ๊ฐ๋ฅํจ.</p>
<ul>
<li><strong>Tubularity Metric</strong>: ํ์ต๋ ๊ถค์ (Trajectory)์ด ๊ตญ์์ ์ผ๋ก ์ํตํ(Tubular) ๊ตฌ์กฐ๋ฅผ ์ ์งํ๋์ง ์ธก์ ํจ. ์ด๋ ๋ชจ๋ธ์ด ๋
ธ์ด์ฆ์ ์ผ๋ง๋ ๊ฐ๊ฑดํ์ง(Robustness)๋ฅผ ๋ํ๋ด๋ ํต์ฌ ์งํ์.</li>
<li><strong>๊ฒ์ฆ ๋ชฉํ</strong>: DIVER-Neuro ๋ชจ๋ธ์ Tubularity Score๊ฐ ๊ธฐ์กด Transformer ๊ธฐ๋ฐ ๋ชจ๋ธ ๋๋น 30 percent ์ด์ ํฅ์๋จ์ ์
์ฆํจ.</li>
</ul>
<h1>์คํ ๋ฐ ๊ฒ์ฆ ๋ฐฉ๋ฒ๋ก (Methodology)</h1>
<h2>๋ฐ์ดํฐ์
๊ตฌ์ถ (Dataset Construction)</h2>
<p>๋ณธ ์ฐ๊ตฌ๋ ๋ ํ๋(Brain)๊ณผ ์ธ๋ถ ํ๊ฒฝ(Vision)์ ๋์์ ํฌํจํ๋ ๋ฉํฐ๋ชจ๋ฌ ๋ฐ์ดํฐ์
์ ํ์ฉํจ.</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>: ๋น์ ํ Izhikevich ๋ด๋ฐ ๋ชจ๋ธ์ ์ฌ์ฉํ์ฌ, Visual Stimulus์ ์๋์ ์์น์ ๋ฐ์ํ๋ ๊ฐ์์ EEG ์์ฑ.</li>
<li><strong>๋ชฉ์ </strong>: Dual Encoder์ ๊ธฐ๋ณธ ๋์ ๋ฐ Latent Alignment ๊ฒ์ฆ.</li>
</ul>
<h3>Main Dataset: BCM-V (Brain-Computer Interface Multimodal Video)</h3>
<ul>
<li><strong>๊ตฌ์ฑ</strong>: ํผํ์๊ฐ 1์ธ์นญ ์์ (Ego-centric) ์์์ ์์ฒญํ ๋ ์ธก์ ๋ 128์ฑ๋ EEG ๋ฐ Eye-tracking ๋ฐ์ดํฐ.</li>
<li><strong>ํน์ง</strong>: ์์ฐ์ค๋ฌ์ด ํ๊ฒฝ์์์ ์๊ฐ-๋ ํ๋ ์๊ด๊ด๊ณ๋ฅผ ํฌํจ.</li>
</ul>
<h2>ํ์ต ์ ๋ต (Training Strategy)</h2>
<h3>Phase 1: Pre-training (Brain-Visual Alignment)</h3>
<ul>
<li><strong>Contrastive Learning (CLIP Style)</strong>: OmniField(Visual)์ Brain Encoder(LTC)์ Latent Vector $z_v, z_b$ ๊ฐ์ ์ฝ์ฌ์ธ ์ ์ฌ๋๋ฅผ ์ต๋ํํจ.</li>
<li><strong>Masked Modeling (MAE Style)</strong>: Brain Encoder์ ์ผ๋ถ ํ์์คํ
์ ๋ง์คํนํ๊ณ ๋ณต์ํ๋๋ก ํ์ตํ์ฌ ์๊ฐ์ ๋ฌธ๋งฅ(Temporal Context) ํ์ต.</li>
</ul>
<h3>Phase 2: Titans Memory Integration</h3>
<ul>
<li><strong>Surprise-based Update</strong>: ์ ๋ ฌ๋ Latent Vector๋ฅผ Titans์ ์
๋ ฅ. ์์ธก ์ค์ฐจ(Prediction Error)๊ฐ ๋์ "๋๋ผ์ด(Surprising)" ์ ๋ณด๋ง Long-term Memory์ ์ ์ฅ.</li>
<li><strong>Free Energy Minimization</strong>: MIRAS์ <strong>Surprise Metric</strong>์ ์ํ์ ์ผ๋ก Friston์ **Free Energy(์์ธก ๋ถํ์ค์ฑ)**์ ๋๋ฑํจ[@friston2010free]. Titans๋ ์ด Free Energy๋ฅผ ์ต์ํํ๋ ๋ฐฉํฅ์ผ๋ก ๋ฉ๋ชจ๋ฆฌ๋ฅผ ๊ฐฑ์ ํจ์ผ๋ก์จ, ์๋ฌผํ์ ํญ์์ฑ(Homeostasis) ์๋ฆฌ๋ฅผ ๊ตฌํํจ.</li>
</ul>
<h3>Brain-Tuning: Semantic Alignment</h3>
<p>DIVER์ fMRI ๋ฐ์ดํฐ๋ฅผ ์ฌ์ฉํ์ฌ Titans์ ์์ ๋ ๋ฒจ ํ์์ ์๋ฏธ๋ก ์ ์ผ๋ก ์ ๋ ฌํฉ๋๋ค [@benara2025braintuning].</p>
<h3>Biological Validation: Gastric Interoception (Murine Data)</h3>
<p>Titans์ ๊ธฐ์ต ๋ฉ์ปค๋์ฆ์ด ์ค์ ์๋ฌผํ์ ํ๋ก์ ์ ์ฌํ๊ฒ ์๋ํ๋์ง ๊ฒ์ฆํ๊ธฐ ์ํด, <strong>๊น์ฑ์ฐ ๊ต์ํ์ ๋๋ฌผ ์คํ ๋ฐ์ดํฐ</strong>๋ฅผ ํ์ฉํฉ๋๋ค [@kim2020neural].</p>
<ul>
<li><strong>Data Source</strong>: ์์ฅ(Mouse)์ <strong>PB Pdyn ๋ด๋ฐ Two-photon Calcium Imaging</strong> ๋ฐ์ดํฐ (์์ฅ ํฝ์ฐฝ ์์ ๋จ์ผ ๋ด๋ฐ ํ์ฑ).</li>
<li><strong>Hypothesis</strong>: Titans์ <strong>Surprise Metric</strong>์ ์์ฅ์ PB Pdyn ๋ด๋ฐ์ด ๋ณด์ฌ์ฃผ๋ <strong>Sustained Activity (Integrator)</strong> ํจํด๊ณผ ๋์ ์๊ด๊ด๊ณ๋ฅผ ๋ณด์ฌ์ผ ํฉ๋๋ค.</li>
<li><strong>Validation</strong>: ์์ฅ ํฝ์ฐฝ(Gastric Distension) ์๋ฎฌ๋ ์ด์
ํ๊ฒฝ์์, Titans์ Memory State ๋ณํ๊ฐ ์ค์ ์ ๊ฒฝ ๋ฐ์ดํฐ์ ์ญ๋์ฑ๊ณผ ์ผ์นํ๋์ง ์ ๋์ ์ผ๋ก ๋ถ์ํฉ๋๋ค. ์ด๋ AI๊ฐ ๋จ์ํ ์ ๋ณด ์ ์ฅ์ด ์๋, ์์ฒด์ ์ ์ฌํ **'๋ด๋ถ ์ํ ์ ๋ถ(Internal State Integration)'**์ ์ํํจ์ ์ฆ๋ช
ํฉ๋๋ค.</li>
</ul>
<h2>์คํ ๋ฐ ๊ฒ์ฆ ๊ณํ (Validation Plan)</h2>
<ul>
<li><strong>Phase 1</strong>: Toy Dataset(Moving MNIST + Synthetic EEG)์ ์ด์ฉํ ๊ฐ๋
์ฆ๋ช
.</li>
<li><strong>Phase 2</strong>: BCM-V(๋-์ปดํจํฐ ์ธํฐํ์ด์ค) ๋ฐ์ดํฐ์
์ ํ์ฉํ ๋๊ท๋ชจ ํ์ต ๋ฐ Tubularity ์ธก์ .</li>
</ul>
<!-- FILE: 03_allostasis.md -->
<h1>๋ด์์ฉ ๊ฐ๊ฐ์ฅ ๋ฐ ์๋ก์คํํฑ ๋ด๋กํธ์ (Allostatic Neuro-Twin)</h1>
<h2>๊ฐ์: ์์ธก์ ์์กด ๋ณธ๋ฅ (Predictive Allostasis)</h2>
<p>๊ธฐ์กด์ AI ์์ด์ ํธ๋ ์ฐ๋์ค ์ฑ์ ๋ฌธ์ ๊ฐ ๋ฐ์ํ ํ ๋ฐ์ํ๋ <strong>Homeostasis(ํญ์์ฑ, Reactive)</strong> ๋ชจ๋ธ์ ๊ทธ์ณค์. ๋ณธ ๊ณผ์ ๋ ์์คํ
์ด ๋ณํํ๋ ๋ฏธ๋ ํ๊ฒฝ์ ์์ธกํ๊ณ ์ฌ์ ์ ์๋์ง๋ฅผ ์กฐ์ ํ๋ <strong>Allostasis(์๋ก์คํ์์ค, Predictive Regulation)</strong> ๊ฐ๋
์ ๋์
ํ์ฌ, ์์ด์ ํธ์ ์์กด์ฑ์ ๊ทน๋ํํจ.</p>
<p><img src="../../05_figures/allostasis_vs_homeostasis_diagram.png" alt="Interoceptive Allostasis vs Homeostasis" /></p>
<h2>์์คํ
๊ตฌ์กฐ: Neuro-Twin Loop</h2>
<p>์ฌ์ฉ์์ ์ ์ฒด ๋ฐ ํ๊ฒฝ ๋ฐ์ดํฐ๋ฅผ ์ค์๊ฐ์ผ๋ก ๋ฏธ๋ฌ๋งํ๋ **"Digital Neuro-Twin"**์ ๊ตฌ์ถํจ.</p>
<h3>Ubiquitous & Visceral Sensing (์
๋ ฅ)</h3>
<ul>
<li><strong>Visceral Interoception</strong>: ๊น์ฑ์ฐ ๊ต์ํ[@kim2020neural]์ ์ฐ๊ตฌ์ ๊ธฐ๋ฐํ์ฌ, ์์ฅ ํฝ์ฐฝ(Gastric Distension)๊ณผ ๊ฐ์ ๋ด๋ถ ์ฅ๊ธฐ ์ ํธ๋ฅผ <strong>NTS $\rightarrow$ PVH $\rightarrow$ aIC</strong> ๊ฒฝ๋ก๋ฅผ ํตํด ์ค์๊ฐ์ผ๋ก ์ถ์ ํจ. ์ด๋ ๋จ์ํ ํฌ๋ง๊ฐ์ ๋์ด, ์์ด์ ํธ์ **๊ธฐ์ด ์๋์ง ์์ค(Basal Energy Level)**์ ์ ์ํ๋ ํต์ฌ ๋ณ์๋ก ์์ฉํจ.</li>
<li><strong>Passive Sensing</strong>: ์ค๋งํธ์์น/๊ธ๋์ค ๋ฑ ์จ์ด๋ฌ๋ธ ๊ธฐ๊ธฐ๋ฅผ ํตํด ์ฌ๋ฐ ๋ณ์ด๋(HRV), ํผ๋ถ ์ ๋๋(EDA) ๋ฑ ์์จ์ ๊ฒฝ๊ณ ์ ํธ๋ฅผ ์๋์ ์ผ๋ก ์์งํจ.</li>
<li><strong>Neural Proxy based on Gastric State</strong>: ์์ฅ ์ํ(Gastric State)๊ฐ ๋์ ๊ฐ์ฑ(Arousal) ๋ฐ ํ๋ ๋ชจ๋(Exploration vs. Exploitation)๋ฅผ ์กฐ์ ํ๋ค๋ ์ด๋ก [@berntson2021neural]์ ์ ์ฉํ์ฌ, ๋ด์ฅ ์ ํธ๋ก๋ถํฐ ๋์ ๊ฑฐ์์ ์ํ๋ฅผ ์ญ์ถ๋ก ํจ.</li>
</ul>
<h3>Energy Landscape Inference (์ฒ๋ฆฌ)</h3>
<p>๋ ์ํ๋ฅผ ๊ณ ์ฐจ์ <strong>"Energy Landscape(์๋์ง ์งํ)"</strong> ์์ ์ขํ๋ก ๋งคํํจ.</p>
<ul>
<li><strong>Pathological Attractor</strong>: ์ฐ์ธ, ๋ถ์, ๋๋ ์์คํ
๊ณผ๋ถํ ์ํ๋ฅผ '๋น ์ ธ๋์ค๊ธฐ ํ๋ ๊น์ ๊ณจ์ง๊ธฐ(Attractor)'๋ก ์ ์ํจ.</li>
<li><strong>Allostatic Load</strong>: ํ์ฌ ์ํ๊ฐ ์ง์๋ ๊ฒฝ์ฐ ์์คํ
์ ๊ฐํด์ง ๋์ ๋ถํ(Load)๋ฅผ ์์ธก ์๋ฎฌ๋ ์ด์
ํจ.</li>
</ul>
<h3>Predictive Nudge (๊ฐ์
)</h3>
<p>์์คํ
์ด๋ ์ฌ์ฉ์๊ฐ ๋ณ๋ฆฌ์ ๋๊ฐ(Attractor)๋ก ์ง์
ํ๊ธฐ ์ , ์ ์ ์ ์ผ๋ก ๊ฐ์
ํจ.</p>
<ul>
<li><strong>Intervention</strong>: ์กฐ๋ช
, ์จ๋ ์กฐ์ ๋๋ ์ฌ์ฉ์์๊ฒ ์๋ฆผ(Nudge)์ ์ ๊ณตํ์ฌ, ์๋์ง ์งํ ์์ '์์ ์ง๋'๋ก ๊ถค์ ์ ์์ ํจ[Psychology Today 2025].</li>
</ul>
<h2>๊ธฐ๋ ํจ๊ณผ</h2>
<p>๋จ์ํ ํฌ์ค์ผ์ด๋ฅผ ๋์ด, ๊ทนํ ํ๊ฒฝ(์ฌ๋, ์ฐ์ฃผ ๋ฑ)์์ ์ธ๊ฐ๊ณผ AI ์์คํ
์ ๊ณต์(Symbiosis)์ ๊ฐ๋ฅ์ผ ํ๋ **"Energy-Efficient Robust Intelligence"**๋ฅผ ๊ตฌํํจ.</p>
<!-- FILE: 04_validation.md -->
<h1>์ด๋ก ์ ํ๋น์ฑ ๋ฐ ์ฑ๋ฅ ๋ชฉํ (Validation)</h1>
<h2>Theoretical Superiority (์ด๋ก ์ ์ฐ์์ฑ)</h2>
<p>๋ณธ ์ ์์๋ ๋จ์ํ ์คํ์ ์๋๊ฐ ์๋, 2024-2025๋
๋ฐํ๋ ์ต์ <strong>SOTA(State-of-the-Art)</strong> ์ฐ๊ตฌ ๊ฒฐ๊ณผ๋ค์ ๊ธฐ๋ฐํ์ฌ ์ค๊ณ๋์๋ค.</p>
<h3>Transformer vs. Titans/SSM in Brain</h3>
<p>๊ธฐ์กด Transformer(ViT)๋ $O(N^2)$์ ์ฐ์ฐ ๋ณต์ก๋๋ก ์ธํด ๊ธด ์๊ณ์ด์ ์์ฒด ์ ํธ๋ฅผ ์ฒ๋ฆฌํ๋ ๋ฐ ํ๊ณ๊ฐ ์๋ค.</p>
<ul>
<li><strong>๊ทผ๊ฑฐ</strong>: <strong>BrainMamba (Wang et al., 2024)</strong> ์ฐ๊ตฌ์ ๋ฐ๋ฅด๋ฉด, SSM ๊ธฐ๋ฐ ๋ชจ๋ธ์ด Transformer ๋๋น <strong>๋ฉ๋ชจ๋ฆฌ ์ฌ์ฉ๋ 50 percent ์ ๊ฐ</strong>, <strong>์ถ๋ก ์๋ 5๋ฐฐ ํฅ์</strong>์ ๋ฌ์ฑํ๋ฉด์๋ EEG ๋ถ๋ฅ ์ ํ๋๋ ๋ ๋์๋ค.</li>
<li><strong>์ ๋ต</strong>: ๋ณธ ์ฐ๊ตฌ๋ BrainMamba์ SSM ์ฝ์ด๋ฅผ **Titans Memory (Behrens et al., 2025)**๋ก ํ์ฅํ์ฌ, ๋จ์ ๋ถ๋ฅ๋ฅผ ๋์ด **"์์ ์ ์ฃผ๊ธฐ ๊ธฐ์ต(Lifelong Learning)"**์ด ๊ฐ๋ฅํ ์ํคํ
์ฒ๋ฅผ ๊ตฌํํ๋ค.</li>
</ul>
<h2>๋ชฉํ ์ฑ๋ฅ ์งํ (Performance Metrics)</h2>
<h3>1. Robustness: Tubularity Score</h3>
<ul>
<li><strong>์ ์</strong>: Neural Manifold๊ฐ ๋
ธ์ด์ฆ(์ธ๋) ์์์๋ ๊ณ ์ ์ ์์ํ์ ๊ตฌ์กฐ(Topology)๋ฅผ ์ ์งํ๋ ์ ๋.</li>
<li><strong>Target</strong>: Baseline (ResNet/ViT) ๋๋น <strong>+30 percent ํฅ์</strong>. (๊ทผ๊ฑฐ: Bertram et al., 2026).</li>
</ul>
<h3>2. Efficiency: Effective Context Length</h3>
<ul>
<li><strong>์ ์</strong>: ์์ค ์์ด ์ญ์ ํ ๊ฐ๋ฅํ ์ต๋ ์ํ์ค ๊ธธ์ด.</li>
<li><strong>Target</strong>: <strong>1M Tokens</strong> (Titans Memory ์ ์ฉ ์). ๊ธฐ์กด RNN(1k), Transformer(8k-32k)๋ฅผ ์๋ํจ.</li>
</ul>
<h3>3. Generalization: Zero-shot Adaptation</h3>
<ul>
<li><strong>์ ์</strong>: ํ์ตํ์ง ์์ ์๋ก์ด ํผํ์(New Subject)์ ๋ํ ์ ์ ๋ฅ๋ ฅ.</li>
<li><strong>Target</strong>: ํ์ต ๋ฐ์ดํฐ ์์ด๋ <strong>Test-time Training</strong>์ ํตํด ์ฆ์ 80 percent ์ด์์ ์ ํ๋ ํ๋ณด.</li>
</ul>
<h2>๊ฒฐ๋ก </h2>
<p>DIVER-Neuro ๋ชจ๋ธ์ <strong>Titans์ ๊ธฐ์ต ๋ฅ๋ ฅ</strong>๊ณผ <strong>SSM์ ํจ์จ์ฑ</strong>์ **์ ์ฒดํ(Embodiment)**๋ ์์ด์ ํธ์ ํตํฉํจ์ผ๋ก์จ, ์ฐจ์ธ๋ AI์ ์๋ก์ด ํ์ค(Standard)์ ์ ์ํ๋ค.</p>
<!-- FILE: 99_red_team_review.md -->
<h1>๐ก๏ธ 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>๐จ 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>๐จ 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>๐จ 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>๐จ 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>