Foundations of Neural Intelligence: μΈλ₯ μ²λ μ 곡νμ μν μ°¨μΈλ μ κ²½κ³Όν μΈμ¬ μμ±
π Live Website: snu-connectome-fellows.vercel.app
"100λ , 1000λ μ μ± μμ§κ³ μΈλ₯μ μ²λ μ 곡νμ ν μ μΈκ³ 0.001%μ ν΅μ¬ μΈμ¬ μμ±"
SNU Connectome Fellows Programμ μμΈλνκ΅ Connectome Labμμ μ΄μνλ νλΆμ μ°κ΅¬ ν λ‘μ°μ νλ‘κ·Έλ¨μ λλ€. λ³Έ νλ‘κ·Έλ¨μ μ κ²½κ³Όν Foundation Model μ°κ΅¬μ μ΅μ μ μμ λ€μ μΈλμ κ³Όνμ 리λλ₯Ό μμ±νλ κ²μ λͺ©νλ‘ ν©λλ€.
- κ³Όνμ μ°κ΅¬: λ¨μν Transformer 곡νμ΄ μλ, μμ± λͺ¨λΈκ³Ό κ²°ν©λ Foundation Model κ°λ°
- λ©ν°λͺ¨λ¬ νμ νμ΅: λ€μν νμμ μ κ²½κ³Όν λ°μ΄ν°λ‘λΆν° 곡ν΅λ λ νμ(representation) λ°κ²¬
- μΈλ₯ 곡ν: λνλ―Όκ΅κ³Ό μΈλ₯μ 100λ , 1000λ λ°μ μ κΈ°μ¬ν ν΅μ¬ μΈμ¬ μ‘μ±
λ³Έ νλ‘κ·Έλ¨μ λ€μμ μ΅μ²¨λ¨ μ°κ΅¬ λ°©ν₯μ μΆκ΅¬ν©λλ€:
| μ°κ΅¬ λΆμΌ | ν΅μ¬ λͺ¨λΈ/λ°©λ²λ‘ | λͺ©ν |
|---|---|---|
| Brain Foundation Models | BrainLM, Brain-JEPA, Brain Harmony | λ νλμ universal representation νμ΅ |
| Generative + Foundation | VAE-JEPA, Diffusion Foundation | λ λ°μ΄ν° μμ± λ° μμΈ‘ ν΅ν© |
| Multimodal Alignment | fMRI-LLM, BrainLLM | μΈμ΄-λ νμ μ λ ¬ |
| Graph Neural Networks | Brain Graph FM | λ μ°κ²°μ± κ·Έλν λͺ¨λΈλ§ |
- BrainLM (ICLR 2024) - fMRI κΈ°λ° Foundation Model
- Brain-JEPA (NeurIPS 2024) - Joint-Embedding Predictive Architecture
- Brain Harmony (NeurIPS 2025) - Multimodal ννν-κΈ°λ₯ ν΅ν©
- Nature 2025 - Foundation model of neural activity predicts response to new stimulus types
| νκ³Ό/μ 곡 | μ°λ 쑰건 |
|---|---|
| μκ³Όλν | μ κ²½κ³Όν/μ μ 건κ°μνκ³Ό κ΄μ¬μ |
| μ κΈ°μ 보곡νλΆ | λ₯λ¬λ/μ νΈμ²λ¦¬ κ²½νμ |
| μ¬λ¦¬νκ³Ό | μΈμ§μ κ²½κ³Όν κ΄μ¬μ |
| μμ μ 곡νλΆ | μ΅ν©μ°κ΅¬ μμμ |
| λμΈμ§κ³Όνκ³Ό | κ³μ°μ κ²½κ³Όν κ΄μ¬μ |
| μ»΄ν¨ν°κ³΅νλΆ | AI/ML μ°κ΅¬ κ²½νμ |
- νμ μ±μ·¨λ (GPA 4.0 μ΄μ κΆμ₯)
- μ°κ΅¬ λκΈ° λ° λΉμ
- νλ‘κ·Έλλ° λ₯λ ₯ (Python, PyTorch)
- μμ΄ λ₯λ ₯ (κ΅μ νλ ₯ νμ)
- κ°μ₯ μ€μ: μΈλ₯μ κΈ°μ¬νκ³ μ νλ μ΄μ κ³Ό ν¬λΆ
λ³Έ νλ‘κ·Έλ¨μ νμ 1λͺ μκ² μ°κ° μ½ 3,620λ§μμ μ§μ ν¬μλ₯Ό μ 곡ν©λλ€.
| κ΅¬λΆ | μ§μ νλͺ© | μ°κ° κΈμ‘ | μμΈ |
|---|---|---|---|
| π΅ νκΈ μ§μ | μ μ°κ΅¬μ₯λ €κΈ | 12,000,000μ | μ 100λ§μ Γ 12κ°μ |
| π ν΄μΈ νλ | ν΄μΈ νν μ§μ | 5,000,000μ | μ° 1ν μ μ‘ (λ±λ‘λΉ, ν곡, μλ°) |
| ν΄μΈ μ°κ΅¬ λ°©λ¬Έ | 5,000,000μ | λ°©ν μ€ λ©ν μ°κ΅¬μ€ λ°©λ¬Έ | |
| π€ AI 리μμ€ | AI API μ¬μ©λ£ | 3,600,000μ | Claude, GPT-5, Gemini λ± (μ 30λ§μ) |
| AI Agent ꡬλ λΉ | 3,600,000μ | Coding Agent 무μ ν μ¬μ© | |
| π νμ΅ μ§μ | λμ/λ Όλ¬Έ ꡬμ | 1,000,000μ | μ°κ΅¬ κ΄λ ¨ μμ ꡬμ |
| π₯οΈ μ₯λΉ ν¬μ | NVIDIA DGX Spark | 6,000,000μ | νμλΉ κ°μΈμ© AI μνΌμ»΄ν¨ν° 1λ |
| μ΄ν© | 36,200,000μ |
μ£Όμ νΉμ§:
- π° μ°κ° νκΈ/μλΉμ€ νν: 3,020λ§μ
- π₯οΈ μ΄κΈ° μ₯λΉ ν¬μ: 600λ§μ (κ°μΈμ© AI μνΌμ»΄ν¨ν°)
- π ν΄μΈ νλ μ§μ: 1,000λ§μ (νν + μ°κ΅¬ λ°©λ¬Έ)
- π€ AI 리μμ€: 720λ§μ (API + Agent ꡬλ )
π‘ μ΄ νλ‘κ·Έλ¨μ νμ 1λͺ λΉ μ°κ° μ½ 3,620λ§μμ μ§μ ν¬μλ₯Ό ν΅ν΄ μ°¨μΈλ μ κ²½κ³Όν μ°κ΅¬ μΈμ¬λ₯Ό μμ±ν©λλ€.
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β Connectome Lab μ»΄ν¨ν
μΈνλΌ β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β β’ NVIDIA DGX Spark (κ°μΈμ© AI μνΌμ»΄ν¨ν°) β
β β’ Lab μλ²: DGX A100, H100 ν΄λ¬μ€ν° β
β β’ ν΄λΌμ°λ: AWS/GCP ν¬λ λ§ μ§μ β
β β’ Coding Agent: Claude Code, Cursor Pro 무μ ν β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
| λ©ν | μμ | μ λ¬Έ λΆμΌ | μν |
|---|---|---|---|
| μ μ μ¬ κ΅μ | BNL (Brookhaven National Lab) | λμμ/μ κ²½κ³Όν | μ°κ΅¬ μλ¬Έ |
| λ°κΈ°ν λ°μ¬ | BNL | κ³μ°μ κ²½κ³Όν | κΈ°μ λ©ν λ§ |
| Uri Hasson | Princeton University | μΈμ΄-λ λͺ¨λΈλ§ | νμ μλ¬Έ |
| Connectome Lab PI | μμΈλνκ΅ | Foundation Models | μ΄κ΄ μ§λ |
| νλ | 보μ |
|---|---|
| μ°κ° μ°κ΅¬ μλ¬Έ (10μκ°+) | $5,000 |
| νμ μ°κ΅¬μ€ λ°©λ¬Έ νΈμ€ν | $3,000 + μλ°/κ΅ν΅λΉ |
| 곡λ λ Όλ¬Έ μ§λ | μ°κ΅¬λΉ μ§μ |
| μ¨λΌμΈ μΈλ―Έλ (2μκ°) | $500 |
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β Annual Program Timeline β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ€
β β
β [3μ] μ λ° λ° μ€λ¦¬μν
μ΄μ
β
β β’ ν λ‘μ° μ λ° (μλ₯ + λ©΄μ ) β
β β’ μ°κ΅¬ λ°©ν₯ μ€μ μν¬μ β
β β’ μ»΄ν¨ν
νκ²½ μ
μ
β
β β
β [4-6μ] νκΈ° μ€ μ°κ΅¬ I β
β β’ μ£Ό 2ν λ© λ―Έν
β
β β’ κ°μΈ μ°κ΅¬ νλ‘μ νΈ μμ β
β β’ μκ° μ§ν λ³΄κ³ β
β β
β [7-8μ] μ¬λ¦ μ§μ€ μ°κ΅¬ β
β β’ ν΄μΈ λ©ν μ°κ΅¬μ€ λ°©λ¬Έ (4μ£Ό) β
β β’ κ΅μ μΈλ¨Έμ€μΏ¨ μ°Έκ° β
β β’ μ§μ€ μ½λ©/μ€ν κΈ°κ° β
β β
β [9-11μ] νκΈ° μ€ μ°κ΅¬ II β
β β’ μ°κ΅¬ κ²°κ³Ό μ 리 β
β β’ λ
Όλ¬Έ μ΄κ³ μμ± β
β β’ κ΅μ νν λ°ν μ€λΉ β
β β
β [12μ] μ°λ§ μ¬ν¬μ§μ β
β β’ ν λ‘μ° μ°κ΅¬ λ°νν β
β β’ μ°¨κΈ° μ°λ κ³ν β
β β’ λ€νΈμνΉ μ΄λ²€νΈ β
β β
β [1-2μ] κ²¨μΈ λ°©ν μ°κ΅¬ β
β β’ μΆκ° ν΄μΈ μ°κ΅¬ κΈ°ν β
β β’ λ
Όλ¬Έ ν¬κ³ μ€λΉ β
β β’ λ€μ νκΈ° μ€λΉ β
β β
βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Connectome Fellows νμ€ κ°λ° νκ²½
# 1. Foundation Models
frameworks = [
"PyTorch 2.5+",
"JAX/Flax",
"Hugging Face Transformers",
"Lightning AI",
]
# 2. Brain-specific Libraries
neuro_tools = [
"nilearn", # fMRI λΆμ
"MNE-Python", # EEG/MEG
"BrainLM", # Foundation Model
"Brain-JEPA", # JEPA ꡬν
"PyTorch Geometric", # Graph Neural Networks
]
# 3. AI Coding Assistants
coding_agents = [
"Claude Code (Opus 4)",
"Cursor Pro (무μ ν)",
"GitHub Copilot Enterprise",
]
# 4. LLM APIs
llm_budget_per_student = {
"Anthropic Claude": "$200/μ",
"OpenAI GPT-5": "$100/μ",
"Google Gemini": "$100/μ",
"DeepSeek R1": "$50/μ",
}λͺ©ν: fMRI, EEG, MEG λ°μ΄ν°λ₯Ό ν΅ν©νλ λ©ν°λͺ¨λ¬ Foundation Model κ°λ°
λ°©λ²λ‘ :
1. JEPA κΈ°λ° self-supervised learning
2. Cross-modal representation alignment
3. Zero-shot transfer to downstream tasks
μμ κ²°κ³Ό:
- Brain state prediction accuracy: 85%+
- Cross-dataset generalization
- NeurIPS/ICLR λ
Όλ¬Έ ν¬κ³
λͺ©ν: λ νλ ν¨ν΄μ μμ±νλ Diffusion-based λͺ¨λΈ κ°λ°
λ°©λ²λ‘ :
1. Latent diffusion for brain signals
2. Conditional generation (stimulus β brain activity)
3. Brain-to-language decoding
μμ κ²°κ³Ό:
- Neural reconstruction quality: SSIM > 0.7
- Language decoding accuracy: 60%+
- Nature Communications λ
Όλ¬Έ λͺ©ν
λͺ©ν: λκ·λͺ¨ μΈμ΄ λͺ¨λΈκ³Ό μΈκ° λ νμμ μ λ ¬ μ°κ΅¬
λ°©λ²λ‘ :
1. Encoding model (LLM β Brain)
2. Representation similarity analysis
3. Shared computational principles λ°κ²¬
λ©ν : Uri Hasson (Princeton)
μμ κ²°κ³Ό:
- Brain-model correlation: r > 0.5
- Insight into language processing
- PNAS/Nature Neuroscience λ
Όλ¬Έ
| κΈ°κ° | μ±κ³Ό λͺ©ν |
|---|---|
| 1λ μ°¨ | κ΅μ νν ν¬μ€ν° λ°ν 1ν |
| 2λ μ°¨ | μ 1μ μ λλ 곡λμ μ λ Όλ¬Έ 1νΈ |
| 3λ μ°¨ | κ΅μ νν°μ΄ νν oral/spotlight |
| μ‘Έμ μ | ν΄μΈ λνμ ν©κ²© λλ μ°κ΅¬μ§ |
- ν΄μΈ λνμ μ§ν μ§μ
- μ°κ΅¬μ/κΈ°μ μ°κ΅¬μ§ μΆμ²
- νμ Connectome Network λ©€λ²μ
| νλͺ© | κΈμ‘ (μ) | λΉκ³ |
|---|---|---|
| νμ μ₯νκΈ | 120,000,000 | 200λ§μ Γ 12κ°μ Γ 5λͺ |
| ν΄μΈ νν/μ¬ν | 50,000,000 | 1,000λ§μ Γ 5λͺ |
| AI μ¬μ© λΉμ© | 30,000,000 | 50λ§μ Γ 12κ°μ Γ 5λͺ |
| λ©ν μ¬λ‘λΉ | 50,000,000 | ν΄μΈ λ©ν 4μΈ |
| μ₯λΉ (DGX Spark) | 100,000,000 | μ΄κΈ° ν¬μ |
| μ΄μλΉ | 20,000,000 | νμ¬, μΈλ―Έλ λ± |
| μλΉλΉ | 30,000,000 | 10% μ¬μ λΆ |
| μ΄κ³ | 400,000,000 | μ½ 4μ΅μ/λ |
- κ΅λ΄ μ§μ: BK21, νμμ°κ΅¬μ§μμ¬μ
- μ λΆ κ³Όμ : NRF, IITP μ°κ΅¬λΉ
- κΈ°μ νμ: μΌμ±, LG, NVIDIA, λ€μ΄λ²
- ν΄μΈ νλ ₯: NIH, NSF 곡λ μ°κ΅¬
- Next.js 14 with App Router and Server Components
- TypeScript for type safety and developer experience
- Tailwind CSS with custom design system
- Framer Motion for smooth animations
- React Hook Form + Zod for robust form validation
- π μμ ν ν/μ μ΄μ€μΈμ΄ μ§μ
- π₯ λ©ν -ν λ‘μ° λ§€μΉ μμ€ν
- π¬ μ°κ΅¬ νλ‘μ νΈ μΆμ μμ€ν
- π κ΄λ¦¬μ λμ보λ λ° λΆμ
- βΏ WCAG 2.1 AA μ κ·Όμ± μ€μ
- π± μμ ν λ°μν λμμΈ
- Fork this repository
- Connect to Vercel
- Set environment variables
- Deploy automatically!
# Clone repository
git clone https://github.com/transconnectome/snu-connectome-fellows.git
cd snu-connectome-fellows/website
# Install dependencies
npm install
# Copy environment file
cp .env.example .env.local
# Run development server
npm run dev- μλ₯ μ μ: μκΈ°μκ°μ, μ°κ΅¬κ³νμ, μ±μ μ¦λͺ μ
- 1μ°¨ μ¬μ¬: μλ₯ νκ°
- 2μ°¨ λ©΄μ : μ°κ΅¬ λ°ν λ° μΈν°λ·°
- μ΅μ’ μ λ°: μ½ 5λͺ λ΄μΈ
- μ΄λ©μΌ: connectome-fellows@snu.ac.kr
- μΉμ¬μ΄νΈ: connectome.snu.ac.kr
- μ°κ΅¬μ€: μμΈλνκ΅ XXXκ΄ XXXνΈ
- Foundation model of neural activity - Nature 2025
- BrainLM Paper - ICLR 2024
- Brain-JEPA - NeurIPS 2024
- Hasson Lab Research - Princeton
- Brain Harmony - NeurIPS 2025
- Deep Learning for Neuroscience
- Computational Neuroscience Course
- Stanford CS330: Multi-Task Learning
π μΈλ₯μ λ―Έλλ₯Ό ν¨κ» λ§λ€μ΄κ° λλ£λ₯Ό μ°Ύμ΅λλ€ π
Last Updated: December 2025
Version: 1.0
Β© 2025 SNU Connectome Lab. All rights reserved.
