neuroconscious-transformer 3.1.4

NeuroConscious Transformer: Next-Generation Neuromorphic Consciousness Architecture

🧠 NeuroConscious Transformer (NCT)

Version: v3.1.3 | 版本: v3.1.3
Author: WENG YONGGANG(翁勇刚) | 邮箱: [e-mail]

---

📦 Installation / 安装

Basic Installation / 基础安装

pip install neuroconscious-transformer

Complete Installation (with Dashboard) / 完整安装（含可视化工具）

pip install neuroconscious-transformer[dashboard]

Verify Installation / 验证安装

python -c "from nct_modules import NCTManager; print('✓ Installation successful')"

Requirements / 依赖要求

• Python 3.9+
• PyTorch 2.0+
• NumPy 1.24+
• SciPy 1.10+

---

🚀 Quick Start / 快速开始

Example 1: Basic Usage / 基础使用

from nct_modules import NCTManager, NCTConfig
import numpy as np

# Create configuration / 创建配置
config = NCTConfig(
    n_heads=8,           # Number of attention heads / 注意力头数
    n_layers=4,          # Number of Transformer layers / Transformer 层数
    d_model=768,         # Representation dimension / 表征维度
    gamma_freq=40.0,     # Gamma wave frequency (Hz) / γ波频率
)

# Initialize manager / 初始化管理器
manager = NCTManager(config)
manager.start()

# Process consciousness cycles / 处理意识周期
for cycle in range(10):
    # Prepare sensory input / 准备感觉输入
    sensory_data = {
        'visual': np.random.randn(28, 28).astype(np.float32),
        'auditory': np.random.randn(10, 10).astype(np.float32),
        'interoceptive': np.random.randn(10).astype(np.float32),
    }
    
    # Process cycle / 处理周期
    state = manager.process_cycle(sensory_data)
    
    # View results / 查看结果
    print(f"Cycle {cycle + 1}:")
    print(f"  Φ Value (Integrated Information): {state.consciousness_metrics.get('phi_value', 0):.3f}")
    print(f"  Free Energy (Prediction Error): {state.self_representation['free_energy']:.4f}")

# Stop / 停止
manager.stop()

Example 2: Multimodal Encoding / 多模态编码

from nct_modules import MultiModalEncoder
import torch

encoder = MultiModalEncoder(
    visual_embed_dim=256,
    audio_embed_dim=256,
    intero_embed_dim=256,
)

# Prepare inputs / 准备输入
visual_input = torch.randn(1, 3, 28, 28)  # RGB image / RGB 图像
audio_input = torch.randn(1, 10, 10)       # Audio spectrogram / 音频频谱
intero_input = torch.randn(1, 10)          # Interoceptive signals / 内感受信号

sensory_tensors = {
    'visual': visual_input,
    'auditory': audio_input,
    'interoceptive': intero_input,
}

# Encode / 编码
embeddings = encoder(sensory_tensors)
print(f"Visual embedding shape: {embeddings['visual'].shape}")  # [1, 768]

Example 3: Φ Value Computation / Φ值计算

from nct_modules import PhiFromAttention
import torch

phi_calc = PhiFromAttention()

# Simulate attention maps / 模拟注意力图
attention_maps = torch.rand(8, 768, 768)  # [heads, seq_len, seq_len]

# Compute Φ value / 计算 Φ值
phi_value = phi_calc.compute_phi(attention_maps)
print(f"Integrated Information Φ: {phi_value:.3f}")
print(f"Consciousness level: {'High' if phi_value > 0.5 else 'Medium' if phi_value > 0.2 else 'Low'}")

Example 4: Transformer-STDP Hybrid Learning / 混合学习

from nct_modules import TransformerSTDP, STDPEvent
import torch

stdp_learner = TransformerSTDP(n_neurons=768, d_model=768)

# Create STDP event / 创建 STDP 事件
event = STDPEvent(
    pre_idx=10,      # Pre-synaptic neuron index / 突触前神经元索引
    post_idx=20,     # Post-synaptic neuron index / 突触后神经元索引
    delta_t=0.015,   # Time difference (seconds) / 时间差
)

# Update synapse / 更新突触
synaptic_update = stdp_learner.update(event)
print(f"Synaptic strength change: Δw = {synaptic_update.delta_w:.6f}")

Example 5: Complete Experiment / 完整实验

"""
NCT Consciousness Computation Experiment
展示：Multimodal Fusion + Φ Monitoring + STDP Learning
"""
from nct_modules import NCTManager, NCTConfig
import numpy as np
import matplotlib.pyplot as plt

# Configuration / 配置
config = NCTConfig(
    n_heads=8,
    n_layers=4,
    d_model=768,
    stdp_learning_rate=0.01,
)

manager = NCTManager(config)
manager.start()

# Record data / 记录数据
phi_values = []
free_energies = []

# Run 100 cycles / 运行 100 个周期
for cycle in range(100):
    # Generate meaningful sensory input (with patterns) / 生成有意义的感觉输入
    visual = np.sin(np.linspace(0, cycle * 0.1, 28 * 28)).reshape(28, 28).astype(np.float32)
    audio = np.cos(np.linspace(0, cycle * 0.05, 10 * 10)).reshape(10, 10).astype(np.float32)
    intero = np.random.randn(10).astype(np.float32) * 0.5
    
    sensory_data = {
        'visual': visual,
        'auditory': audio,
        'interoceptive': intero,
    }
    
    # Process / 处理
    state = manager.process_cycle(sensory_data)
    
    # Record / 记录
    phi_values.append(state.consciousness_metrics.get('phi_value', 0))
    free_energies.append(state.self_representation['free_energy'])
    
    # Print every 10 cycles / 每 10 个周期打印
    if (cycle + 1) % 10 == 0:
        print(f"Cycle {cycle + 1}/100 | Φ={phi_values[-1]:.3f} | FE={free_energies[-1]:.4f}")

manager.stop()

# Visualization / 可视化
plt.figure(figsize=(12, 5))

plt.subplot(1, 2, 1)
plt.plot(phi_values)
plt.title('Φ Value (Integrated Information)')
plt.xlabel('Cycle')
plt.ylabel('Φ')
plt.grid(True, alpha=0.3)

plt.subplot(1, 2, 2)
plt.plot(free_energies)
plt.title('Free Energy (Prediction Error)')
plt.xlabel('Cycle')
plt.ylabel('FE')
plt.grid(True, alpha=0.3)

plt.tight_layout()
plt.savefig('nct_experiment_results.png', dpi=300)
plt.show()

print("\n✅ Experiment complete! Results saved to nct_experiment_results.png")

---

🔧 Core API Reference / 核心 API 速查

NCTConfig - Configuration Class / 配置类

from nct_modules import NCTConfig

config = NCTConfig(
    # Core parameters / 核心参数
    n_heads=8,                    # Attention heads (workspace capacity) / 注意力头数
    n_layers=4,                   # Transformer layers / Transformer 层数
    d_model=768,                  # Representation dimension / 表征维度
    dim_ff=3072,                  # Feed-forward network dimension / 前馈网络维度
    
    # Neuroscience parameters / 神经科学参数
    gamma_freq=40.0,              # Gamma wave frequency (Hz) / γ波频率
    stdp_learning_rate=0.01,      # STDP learning rate / STDP 学习率
    attention_modulation_lambda=0.1,  # Attention modulation coefficient / 注意力调制系数
)

NCTManager - Manager Class / 管理器类

from nct_modules import NCTManager

# Initialize / 初始化
manager = NCTManager(config)

# Start system / 启动系统
manager.start()

# Process one consciousness cycle / 处理一个意识周期
state = manager.process_cycle(sensory_data)

# Stop system / 停止系统
manager.stop()

# Get statistics / 获取统计信息
stats = manager.get_stats()

# Save model / 保存模型
torch.save(manager.state_dict(), 'model.pth')

# Load model / 加载模型
manager.load_state_dict(torch.load('model.pth'))

ConsciousnessState - State Object / 状态对象

# Access consciousness metrics / 访问意识度量
state.awareness_level              # Consciousness level (low/moderate/high) / 意识水平
state.consciousness_metrics        # Metrics dictionary (includes Φ value) / 意识度量字典
state.self_representation          # Self-representation (free energy, confidence) / 自我表征
state.workspace_content            # Global workspace content / 全局工作空间内容

# Example usage / 示例用法
print(f"Consciousness level: {state.awareness_level}")
print(f"Φ value: {state.consciousness_metrics['phi_value']:.3f}")
print(f"Free energy: {state.self_representation['free_energy']:.4f}")

Other Core Components / 其他核心组件

MultiModalEncoder / 多模态编码器

from nct_modules import MultiModalEncoder

encoder = MultiModalEncoder(
    visual_embed_dim=256,
    audio_embed_dim=256,
    intero_embed_dim=256,
)

embeddings = encoder(sensory_tensors)

PhiFromAttention / Φ值计算器

from nct_modules import PhiFromAttention

phi_calc = PhiFromAttention()
phi_value = phi_calc.compute_phi(attention_maps)

TransformerSTDP / 混合学习

from nct_modules import TransformerSTDP, STDPEvent

stdp = TransformerSTDP(n_neurons=768, d_model=768)
event = STDPEvent(pre_idx=10, post_idx=20, delta_t=0.015)
update = stdp.update(event)

NCTWorkspace / 全局工作空间

from nct_modules import NCTWorkspace

workspace = NCTWorkspace(n_heads=8, d_model=768)
global_content = workspace(attention_maps, query)

---

🎨 Advanced Features / 高级功能

GPU Acceleration / GPU 加速

import torch
torch.set_default_device('cuda')  # Automatically use GPU / 自动使用 GPU

# All computations will now run on GPU / 所有计算将在 GPU 上运行
config = NCTConfig()
manager = NCTManager(config)

Custom Sensory Inputs / 自定义感觉输入

# Supports arbitrary shapes (automatically adapts) / 支持任意形状（自动适配）
sensory_data = {
    'visual': your_image,           # [H, W] or [C, H, W]
    'auditory': your_audio,         # [T, F]
    'interoceptive': your_signals,  # [N]
}

Model Checkpointing / 模型检查点

import torch
from nct_modules import NCTManager, NCTConfig

# Save checkpoint / 保存检查点
checkpoint = {
    'config': config.to_dict(),
    'model_state': manager.state_dict(),
    'metrics': manager.get_stats(),
    'cycle': current_cycle,
}
torch.save(checkpoint, 'nct_checkpoint.pth')

# Load checkpoint / 加载检查点
loaded = torch.load('nct_checkpoint.pth')
loaded_config = NCTConfig.from_dict(loaded['config'])
loaded_manager = NCTManager(loaded_config)
loaded_manager.load_state_dict(loaded['model_state'])

print(f"✓ Checkpoint loaded at cycle {loaded['cycle']}")

Preset Configurations / 预设配置

# Large-scale configuration (research) / 大规模配置（研究用）
large_config = NCTConfig(
    n_heads=12,
    n_layers=6,
    d_model=1024,
    dim_ff=4096,
)

# Lightweight configuration (real-time) / 轻量级配置（实时应用）
small_config = NCTConfig(
    n_heads=4,
    n_layers=2,
    d_model=256,
    gamma_freq=30.0,
)

# Temporal learning configuration / 时序关联学习配置
temporal_config = NCTConfig(
    n_heads=8,
    n_layers=4,
    d_model=512,
    stdp_learning_rate=0.05,
    attention_modulation_lambda=0.2,
)

---

📊 Common Issues / 常见问题

Q1: How to improve Φ value? / 如何提高 Φ值？

• Increase d_model (representation dimension) / 增加 d_model（表征维度）
• Increase n_heads (attention head count) / 增加 n_heads（注意力头数）
• Use more meaningful sensory inputs (patterned data) / 使用更有意义的感觉输入

Q2: How to use custom modalities? / 如何使用自定义模态？

# Add any modality you want / 添加任意模态
sensory_data = {
    'custom_modality_1': tensor1,
    'custom_modality_2': tensor2,
}

# The encoder will automatically adapt / 编码器会自动适配

Q3: Dashboard not launching? / Dashboard 无法启动？

# Install dashboard dependencies / 安装 Dashboard 依赖
pip install neuroconscious-transformer[dashboard]

# Or manually install / 或手动安装
pip install streamlit plotly pandas

# Launch / 启动
nct-dashboard

---

📚 Further Reading / 延伸阅读

Technical Blog Series / 技术博客系列

• Consciousness: From Philosophy to Engineering
• [Attention as Global Workspace](https://github.com/winclaw/NCT/tree/main/docs/csdn_blog/02_Attention 如何成为全局工作空间_Miller 定律的深度学习诠释.md)
• Transformer-STDP Integration
• [Predictive Coding & Free Energy](https://github.com/winclaw/NCT/tree/main/docs/csdn_blog/04_预测编码=Decoder 训练_Friston 自由能的 Transformer 实现.md)

Academic Paper / 学术论文

• arXiv preprint: NeuroConscious Transformer (coming soon)
• Technical report with full experiments and ablation studies

GitHub Repository / GitHub 仓库

For source code, examples, and development documentation:
https://github.com/winclaw/NCT

---

👨‍💻 Author & License / 作者与许可

Author / 作者: WENG YONGGANG (翁勇刚)
Affiliation / 机构: Faculty of Computer Science and Information Technology, Universiti Teknologi Malaysia
Email / 邮箱: [e-mail]

License / 许可证: MIT License

---

🙏 Acknowledgments / 致谢

This project integrates insights from:

• Transformer architecture (Vaswani et al., 2017)
• STDP mechanisms (Bi & Poo, 1998)
• Predictive Coding theory (Rao & Ballard, 1999)
• Integrated Information Theory (Tononi et al.)
• Global Workspace Theory (Baars, Dehaene)

本项目融合了以下理论洞见：

• Transformer 架构（Vaswani 等，2017）
• STDP 机制（Bi & Poo, 1998）
• 预测编码理论（Rao & Ballard, 1999）
• 整合信息论（Tononi 等）
• 全局工作空间理论（Baars, Dehaene）

---

📬 Contact / 联系

• GitHub Issues: https://github.com/winclaw/NCT/issues
• Email: [e-mail]
• PyPI: https://pypi.org/project/neuroconscious-transformer/

Welcome to join the NeuroConscious community! 🧠✨
欢迎加入 NeuroConscious 社区！🧠✨