Neuromodulated-Tower-System

Biologically-Inspired Multi-Tower Architecture with Neuromodulator-Gated Integration

A PyTorch-based implementation of a recursive self-improvement system inspired by Hansen et al.'s (2024) brainstem-cortex connectivity findings. The current iteration emphasizes stability and interpretability through normalized encoders, hormone-banded affective signals, and receptor-aware routing in the neuromodulator gate.

Architecture Overview

5 Specialized Processing Towers

Tower 1: Social-Memory

• Autobiographical and episodic memory processing
• Social cognition and theory of mind
• Long-term pattern storage via EWC (Elastic Weight Consolidation)

Tower 2: Working-Memory & Cognitive Control

• Dynamic task management
• Cognitive state monitoring
• Meta-learning policy (learns when and how to use other towers)

Tower 3: Affective Processing

• 3-hormone neuromodulatory system (Dopamine, Serotonin, Cortisol)
• Emotional state representation with LayerNorm
• Intrinsic drive head that keeps affective latent expressive and gradients healthy
• Clamped hormone bands to avoid saturation

Tower 4: Sensorimotor Integration

• Perception (vision, proprioception)
• Action decoding (dual-head: what & where/how)
• Sensory-motor binding

Tower 5: Motor Coordination & Sequencing

• Complex behavioral sequencing
• Planning and trajectory generation
• Execution of motor programs

Neuromodulator Gating Layer

Based on PET imaging findings from Hansen et al. (2024):

• 18-receptor dynamic routing system
• 3 main neurotransmitter pathways: NET (norepinephrine), DAT (dopamine), 5-HTT (serotonin)
• Learnable receptor sensitivities that scale pathway influence
• Baseline router to stabilize routing under noisy hormones
• Context-dependent modulation: Tower outputs weighted by current hormonal state and task context
• Biologically-validated connectivity patterns: Unimodal ↔ Transmodal hierarchy

Mirror Tower (Self-Reflective Refinement)

• EMA reflection state blends prior refined latents with current integrated signal
• Gated residual update pushes latents toward self-consistent directions while respecting safety floors
• Stability-first normalization keeps the reflection loop numerically well-behaved before cortical reasoning

Recursive Self-Improvement Loop

1. Parallel Tower Processing: 5 towers process independently
2. NT-Gated Integration: Neuromodulator gate combines outputs
3. Cortical Reasoning: H-module style planning (inspired by Sapient HRM)
4. Action Selection: L-module execution with meta-learning
5. Feedback Loop: Gate weights updated via meta-cognition module

Installation

git clone https://github.com/sunghunkwag/Neuromodulated-Tower-System.git
cd Neuromodulated-Tower-System
pip install -r requirements.txt

Quick Start

import torch
from src.system import FiveTowerSystem

# Initialize system
system = FiveTowerSystem(
    latent_dim=128,
    device='cuda' if torch.cuda.is_available() else 'cpu'
)

# Process input state
state = torch.randn(1, 256)  # Batch size 1, 256-dim state
action, debug = system(state)

print(f"Action shape: {action.shape}")
print(f"NT gate weights: {debug['nt_weights']}")
print(f"Mirror gate range: [{debug['mirror']['gate'].min().item():.4f}, {debug['mirror']['gate'].max().item():.4f}]")

The neuromodulator gate supports stability controls for more reliable gradients on small batches:

system = FiveTowerSystem(
    latent_dim=128,
    gate_temperature=0.7,       # softens routing logits to avoid overly sharp tower selection
    min_pathway_share=0.05,     # guarantees a small weight for each tower to keep signals flowing
    device='cuda' if torch.cuda.is_available() else 'cpu'
)

Core Features

✅ 5-Tower Parallel Processing: Specialized cognitive modules
✅ Neurotransmitter-Gated Integration: Context-dependent routing
✅ Recursive Meta-Learning: Self-improving gate weights
✅ Biologically-Plausible: Grounded in neuroscience (Hansen et al., 2024)
✅ PyTorch Native: Full GPU support and autograd compatibility
✅ Modular Design: Each tower independently trainable

Project Structure

Neuromodulated-Tower-System/
├── README.md
├── requirements.txt
├── src/
│   ├── __init__.py
│   ├── neuromodulator_gate.py
│   ├── system.py
│   ├── towers/
│   │   ├── __init__.py
│   │   ├── tower_base.py
│   │   ├── tower1_social_memory.py
│   │   ├── tower2_working_memory.py
│   │   ├── tower3_affective.py
│   │   ├── tower4_sensorimotor.py
│   │   └── tower5_motor_coordination.py
│   └── training.py
└── test_validation.py

Key Papers & References

1. Hansen et al. (2024) - "Brainstem-Cortex Connectivity and Hierarchical Cognition"
   Nature Neuroscience - Found 5-community structure in brainstem nuclei

2. Sutskever (2024) - Digital Brainstem Concept
   SSI Research Direction

3. Sapient (2024) - HRM (Hierarchical Reasoning Model)
   ArXiv - H-module + L-module dual processing

4. Kahneman (2011) - Thinking Fast and Slow
   Dual-Process Theory Foundation

Training & Evaluation

Basic Training Loop

from src.training import train_epoch
from src.system import FiveTowerSystem
import torch
from torch.optim import Adam

system = FiveTowerSystem(latent_dim=128, device='cuda')
optimizer = Adam(system.parameters(), lr=1e-3)

for epoch in range(100):
    loss = train_epoch(system, optimizer, train_loader)
    print(f"Epoch {epoch}: Loss={loss:.4f}")

Testing

python -m pytest -q

The default test runner executes test_validation.py to ensure the towers, neuromodulator gate, and integration pipeline produce valid shapes and stable hormone-aware routing.

SCIG (Self-Contracting Improvement Graph) Demo

The repository also includes a standalone scig_demo.py script that showcases a recursive improvement loop with adversarial test forging and self-tuning patch policies. To run the demo:

python scig_demo.py

The script prints the best discovered expression in its DSL, along with operator statistics and the adaptive focus strength used by the evaluator.

Contributing

Contributions welcome! Please:

1. Fork repository
2. Create feature branch (git checkout -b feature/amazing-feature)
3. Commit changes (git commit -m 'Add amazing feature')
4. Push to branch (git push origin feature/amazing-feature)
5. Open Pull Request

License

MIT License - see LICENSE file for details

Citation

If you use this in research, please cite:

@software{kwag2025neuromodulated,
  author = {Kwag, Sunghun},
  title = {Neuromodulated-Tower-System: A Biologically-Inspired Multi-Tower Architecture},
  year = {2025},
  url = {https://github.com/sunghunkwag/Neuromodulated-Tower-System}
}

Author

Sunghun Kwag - Independent AI Research
GitHub: @sunghunkwag

---

Status: 🚀 Active Development
Last Updated: December 2025