本论文提出利用 MICrONS(Machine Intelligence from Cortical Networks)功能连接组数据集构建生物合理的循环神经网络(RNN)。通过整合皮层几何结构、神经连接和功能关系三个维度的生物约束,作为强归纳偏置指导网络学习,显著提升认知决策任务性能并涌现出类脑网络拓扑特性。
Scanned 9/11/2026
Install to Claude Code
npx -y skills add hiyenwong/ai_collection --skill cortical-geometry-rnn-inductive-biases --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Cortical Geometry Rnn Inductive Biases?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/hiyenwong-cortical-geometry-rnn-inductive-biases-ai-collection)More formats (shields.io, HTML) on the badges page.
---
name: cortical-geometry-rnn-inductive-biases
category: neuroscience
tags:
- rna
- computational-neuroscience
- inductive-biases
- functional-connectomics
- spatial-embedding
- MICrONS
- brain-inspired-ai
created: 2026-06-29
paper: arXiv:2606.14975
paper_title: "Harnessing cortical geometry, wiring, and function as inductive biases for recurrent neural networks"
authors:
- Mo Shakiba
- Rana Rokni
- Mohammad Mohammadi
- Nima Dehghani
institution: MIT, Neuromatch
abstract: "Biologically-grounded RNN construction using MICrONS functional connectomics dataset. Combines neuronal spatial coordinates, anatomical connectivity, and functional relationships from ~12,000 excitatory neurons to initialize recurrent weights and impose communication-aware spatial constraints. Networks constrained by cortical structure and function consistently outperform baseline models across cognitive tasks, developing low-entropy, modular, small-world organization."
key_concepts:
- functional-connectomics-derived-weight-initialization
- spatial-embedding-with-real-neuronal-coordinates
- communicability-based-regularization
- MICrONS-multimodal-dataset
- cortical-inductive-biases
- emergent-small-world-modularity
---
# Cortical Geometry, Wiring, and Function as RNN Inductive Biases
## 概述
本论文提出利用 MICrONS(Machine Intelligence from Cortical Networks)功能连接组数据集构建生物合理的循环神经网络(RNN)。通过整合皮层几何结构、神经连接和功能关系三个维度的生物约束,作为强归纳偏置指导网络学习,显著提升认知决策任务性能并涌现出类脑网络拓扑特性。
## 核心方法论
### 1. 多模态数据源(MICrONS 数据集)
- **结构连接**:来自电子显微镜(EM)重建的 ~12,000 个兴奋性神经元解剖连接
- **功能活动**:双光子钙成像获得的神经元活动数据
- **空间坐标**:同一动物中配准的神经元三维空间位置
### 2. 三层归纳偏置架构
#### 2.1 功能初始化(Weight Initialization - W*)
**机制**:从神经元功能相关性矩阵计算初始权重
- 使用 Pearson 相关系数、STTC(Spike Time Tiling Coefficient)或精度矩阵
- 构建生物合理的权重矩阵 W_bio
- 保留功能连接的统计结构而非精确映射
**效果**:
- 提供最大性能增益(准确率提升 10-30%)
- 在纯正权重约束下防止模型崩溃
- 关键发现:W*(精确映射)与 W!(置换映射)性能无显著差异,说明优势来自权重统计结构而非精确位置对应
#### 2.2 真实空间嵌入(Spatial Embedding - D*)
**机制**:使用真实神经元坐标而非人工网格
- 计算神经元对之间的欧氏距离矩阵 D*
- 应用距离衰减正则化:λ‖W ⊙ D*‖
- 约束连接强度随空间距离衰减
**效果**:
- 提供稳健的次要性能提升
- 引导网络向低熵、模块化组织演化
- 真实空间布局作为有意义的计算先验
#### 2.3 可通信性正则化(Communicability - C)
**机制**:基于图论的可通信性度量
- 定义:C = (I - αW)^{-1} - I,衡量节点间多路径信息流
- 正则化项:λ‖W* ⊙ D* ⊙ C‖
- 变体:使用 EMD(Earth Mover's Distance)匹配经验与人工可通信性分布
**效果**:
- 效果较微妙,依赖任务上下文
- 与真实空间嵌入结合时最显著
- 促进小世界特性涌现
### 3. 模型变体系统
论文测试 11 种模型变体,系统解耦各组件贡献:
| 模型 | 功能初始化 | 空间嵌入 | 可通信性 | 性能 |
|------|-----------|---------|---------|------|
| W*D*C | ✓ | 真实 | ✓ | 最优 |
| W*D*C* | ✓ | 真实 | EMD | 次优 |
| W!D*C | 置换 | 真实 | ✓ | 强 |
| WD*C | ✗ | 真实 | ✓ | 中等 |
| W | ✗ | 无 | ✗ | 基线 |
### 4. 认知任务范式
1. **单选择推理**:整合目标和选择刺激,延迟后决策
2. **知觉决策**:从噪声刺激中识别主导方向
3. **Go/NoGo**:响应或抑制反应的二元决策
## 关键发现
### 性能层级
1. **功能初始化**(最大增益)
2. **真实空间嵌入**(稳健次要增益)
3. **可通信性正则化**(任务依赖)
### 涌现网络特性
- **低熵**:高度组织化的权重结构(熵 ~0.6-1.5 vs 基线 ~3-6)
- **模块化**:Q ~0.4-0.5(强社区结构)
- **小世界性**:σ ~1.5-2.5(高聚类 + 短路径)
- **解 assortativity**:r < 0(hub-外围组织)
### 鲁棒性验证
- 精确矩阵置换(W!)保留性能 → 统计结构是关键
- 精度矩阵替代相关矩阵 → 性能相当
- 跨会话/扫描/视野一致性 → 泛化性强
## 实现指南
### 数据准备
```python
import numpy as np
from scipy.spatial.distance import pdist, squareform
# 1. 功能连接矩阵(从钙成像数据计算)
def compute_functional_connectivity(calcium_traces, method='correlation'):
"""
calcium_traces: shape (n_neurons, n_timepoints)
"""
if method == 'correlation':
W_bio = np.corrcoef(calcium_traces)
elif method == 'sttc':
# 实现 STTC 计算
pass
elif method == 'precision':
cov = np.cov(calcium_traces)
W_bio = np.linalg.inv(cov)
# 阈值化稀疏连接
threshold = np.percentile(np.abs(W_bio), 80)
W_bio[np.abs(W_bio) < threshold] = 0
return W_bio
# 2. 距离矩阵(从空间坐标计算)
def compute_distance_matrix(coordinates):
"""
coordinates: shape (n_neurons, 3)
"""
D = squareform(pdist(coordinates, 'euclidean'))
return D
# 3. 可通信性矩阵
def compute_communicability(W, alpha=0.5):
"""
W: 权重矩阵(已归一化)
alpha: 衰减参数
"""
n = W.shape[0]
C = np.linalg.inv(np.eye(n) - alpha * W) - np.eye(n)
return C
```
### RNN 构建
```python
import torch
import torch.nn as nn
class CorticalConstrainedRNN(nn.Module):
def __init__(self, input_size, hidden_size, output_size,
W_bio=None, D_real=None, C_bio=None,
lambda_dist=0.01, lambda_comm=0.01):
super().__init__()
self.hidden_size = hidden_size
# 功能初始化
if W_bio is not None:
self.weight_hh = nn.Parameter(torch.tensor(W_bio, dtype=torch.float32))
else:
self.weight_hh = nn.Parameter(torch.randn(hidden_size, hidden_size))
self.weight_ih = nn.Parameter(torch.randn(4 * hidden_size, input_size))
self.weight_ho = nn.Parameter(torch.randn(output_size, hidden_size))
self.bias_hh = nn.Parameter(torch.zeros(4 * hidden_size))
# 空间约束
self.D_real = torch.tensor(D_real, dtype=torch.float32) if D_real is not None else None
self.C_bio = torch.tensor(C_bio, dtype=torch.float32) if C_bio is not None else None
self.lambda_dist = lambda_dist
self.lambda_comm = lambda_comm
def forward(self, x):
# LSTM 风格实现
batch_size = x.size(0)
h = torch.zeros(batch_size, self.hidden_size, device=x.device)
c = torch.zeros(batch_size, self.hidden_size, device=x.device)
for t in range(x.size(1)):
gates = torch.mm(x[:, t, :], self.weight_ih.T) + \
torch.mm(h, self.weight_hh.T) + self.bias_hh
i, f, g, o = torch.chunk(gates, 4, dim=1)
i = torch.sigmoid(i)
f = torch.sigmoid(f)
g = torch.tanh(g)
o = torch.sigmoid(o)
c = f * c + i * g
h = o * torch.tanh(c)
output = torch.mm(h, self.weight_ho.T)
return output
def compute_regularization(self):
"""计算空间和可通信性正则化"""
reg = 0.0
# 距离正则化
if self.D_real is not None:
dist_reg = torch.norm(self.weight_hh * self.D_real.to(self.weight_hh.device))
reg += self.lambda_dist * dist_reg
# 可通信性正则化
if self.C_bio is not None:
comm_reg = torch.norm(self.weight_hh * self.C_bio.to(self.weight_hh.device))
reg += self.lambda_comm * comm_reg
return reg
```
### 训练循环
```python
def train_cortical_rnn(model, train_loader, epochs=100, lr=0.001):
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
criterion = nn.CrossEntropyLoss()
for epoch in range(epochs):
total_loss = 0
for batch_x, batch_y in train_loader:
optimizer.zero_grad()
outputs = model(batch_x)
task_loss = criterion(outputs, batch_y)
# 添加生物正则化
bio_reg = model.compute_regularization()
loss = task_loss + bio_reg
loss.backward()
optimizer.step()
total_loss += loss.item()
if (epoch + 1) % 10 == 0:
print(f"Epoch {epoch+1}/{epochs}, Loss: {total_loss/len(train_loader):.4f}")
```
## 关键洞察
1. **功能 > 结构 > 正则化**:归纳偏置的层级效应,功能初始化最关键
2. **统计结构 > 精确映射**:权重的整体分布特性比神经元-权重一一对应更重要
3. **多模态约束协同**:三层约束组合产生涌现特性,而非简单叠加
4. **生物合理性与性能统一**:类脑拓扑特性(小世界、模块化)与高性能共存
## 局限性与展望
**局限**:
- 仅测试视觉皮层数据,泛化性待验证
- 计算成本较高(~12k 神经元矩阵运算)
- 未探索抑制性神经元贡献
**展望**:
- 扩展到其他脑区(海马、前额叶)
- 整合抑制性连接和神经调质
- 应用于更大规模任务(语言、推理)
- 与 Transformer 架构结合
## 相关资源
- **代码仓库**:https://github.com/neurovium/CorticalBlueprintRNN
- **数据集**:MICrONS (https://www.micronsproject.org/)
- **相关论文**:
- Turner et al. (2020) MICrONS multiscale reconstruction
- Bae et al. (2025) Functional connectomics in mouse visual cortex
- Bullmore & Sporns (2012) Economy of brain network organization
## 使用场景
- 构建生物合理的 RNN 用于认知建模
- 测试特定脑区连接对计算的影响
- 研究网络拓扑与功能的因果关系
- 开发类脑 AI 架构
## 注意事项
- 需要 MICrONS 数据访问权限
- 大规模矩阵运算需要 GPU 加速
- 超参数(λ_dist, λ_comm)需任务特异性调优
- 功能连接方法选择影响结果(correlation vs precision)
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!