cornn-convex-rnn-optimization

---
name: cornn-convex-rnn-optimization
description: CORNN凸优化递归神经网络方法论。将RNN训练转化为凸优化问题，训练速度比传统方法快100倍，支持百万参数RNN在标准计算机上亚分钟级训练。适用于大规模神经记录实时建模、神经动力学推断、吸引子结构恢复。触发词：RNN训练、凸优化、神经动力学、实时建模、数据约束RNN、convex optimization、recurrent neural network、neural dynamics inference。
user-invocable: true
---

# CORNN: Convex Optimization of Recurrent Neural Networks

凸优化递归神经网络快速神经动力学推断

## 核心方法论

**来源：** arXiv:2311.10200 (NeurIPS 2023)
**效用：** 0.92

### 核心创新

将 RNN 训练转化为凸优化问题：
- 训练速度比传统方法快 **100 倍**
- 百万参数 RNN 在标准计算机上 **亚分钟级训练**
- 支持实时网络重建

### 实现框架

```python
import numpy as np
import cvxpy as cp
from typing import Tuple

class CORNNTrainer:
    """
    CORNN: 凸优化 RNN 训练器
    
    核心思想：将非凸 RNN 训练转化为凸优化问题
    """
    
    def __init__(
        self,
        n_neurons: int,
        hidden_dim: int = 64,
        regularization: float = 0.01
    ):
        self.n_neurons = n_neurons
        self.hidden_dim = hidden_dim
        self.reg = regularization
    
    def train(
        self,
        neural_data: np.ndarray,
        max_iter: int = 1000
    ) -> Tuple[np.ndarray, np.ndarray]:
        """
        凸优化训练
        
        参数:
            neural_data: 神经数据 (n_timepoints, n_neurons)
            max_iter: 最大迭代次数
            
        返回:
            W_rec: 递归权重
            W_in: 输入权重
        """
        T, N = neural_data.shape
        
        # 定义优化变量
        W_rec = cp.Variable((self.hidden_dim, self.hidden_dim))
        W_in = cp.Variable((self.hidden_dim, N))
        h = cp.Variable((T, self.hidden_dim))
        
        # 构建损失函数
        loss = 0
        for t in range(1, T):
            # RNN 动力学约束
            h_pred = W_rec @ h[t-1] + W_in @ neural_data[t-1]
            loss += cp.sum_squares(h[t] - cp.tanh(h_pred))
        
        # 正则化
        loss += self.reg * (cp.norm(W_rec, 'fro') + cp.norm(W_in, 'fro'))
        
        # 求解凸优化问题
        problem = cp.Problem(cp.Minimize(loss))
        problem.solve(max_iter=max_iter)
        
        return W_rec.value, W_in.value
    
    def infer_dynamics(
        self,
        neural_data: np.ndarray,
        W_rec: np.ndarray,
        W_in: np.ndarray
    ) -> np.ndarray:
        """
        推断神经动力学
        
        返回:
            隐藏状态轨迹
        """
        T = neural_data.shape[0]
        h = np.zeros((T, self.hidden_dim))
        
        for t in range(1, T):
            h[t] = np.tanh(W_rec @ h[t-1] + W_in @ neural_data[t-1])
        
        return h


def cornn_train(
    neural_data: np.ndarray,
    hidden_dim: int = 64,
    regularization: float = 0.01
) -> dict:
    """
    CORNN 训练接口
    
    参数:
        neural_data: 神经数据
        hidden_dim: 隐藏维度
        regularization: 正则化系数
        
    返回:
        训练结果
    """
    trainer = CORNNTrainer(
        n_neurons=neural_data.shape[1],
        hidden_dim=hidden_dim,
        regularization=regularization
    )
    
    W_rec, W_in = trainer.train(neural_data)
    dynamics = trainer.infer_dynamics(neural_data, W_rec, W_in)
    
    return {
        'W_rec': W_rec,
        'W_in': W_in,
        'dynamics': dynamics
    }
```

## 应用场景

1. 大规模神经记录实时建模
2. 神经动力学推断
3. 吸引子结构恢复

## 关键优势

| 指标 | 传统方法 | CORNN |
|------|---------|-------|
| 训练速度 | 基准 | **快 100 倍** |
| 参数规模 | 有限 | **百万级** |
| 实时性 | 离线 | **准实时** |

## Activation Keywords
- RNN训练
- 凸优化
- 神经动力学
- 实时建模
- 数据约束RNN
- convex optimization
- recurrent neural network
- neural dynamics inference

## Tools Used
- numpy
- cvxpy

## Instructions for Agents
1. 准备神经数据：确保数据格式为 (时间步, 神经元)
2. 配置隐藏维度和正则化参数
3. 调用 cornn_train 函数进行训练
4. 分析返回的权重矩阵和动力学推断结果
5. 验证训练速度和准确性

## Examples
```python
# 使用示例
from cornn_convex_rnn_optimization import cornn_train
import numpy as np

# 1. 生成模拟神经数据
T, n_neurons = 1000, 50
neural_data = np.random.randn(T, n_neurons)

# 2. 训练 RNN
result = cornn_train(
    neural_data,
    hidden_dim=32,
    regularization=0.01
)

# 3. 查看结果
print(f"递归权重: {result['W_rec'].shape}")
print(f"输入权重: {result['W_in'].shape}")
print(f"动力学轨迹: {result['dynamics'].shape}")
```

## 参考文献
- Dinc, F., et al. (2023). "Convex optimization of recurrent neural networks for rapid inference of neural dynamics" NeurIPS 2023