Free SKILL.md scraped from GitHub. Clone the repo or copy the file directly into your Claude Code skills directory.
npx versuz@latest install hiyenwong-ai-collection-collection-skills-autoregressive-flow-matching-neuralgit clone https://github.com/hiyenwong/ai_collection.gitcp ai_collection/SKILL.MD ~/.claude/skills/hiyenwong-ai-collection-collection-skills-autoregressive-flow-matching-neural/SKILL.md---
name: autoregressive-flow-matching-neural
description: "Autoregressive Flow Matching (AFM) framework for probabilistic prediction of neural dynamics from multimodal sensory input, enabling conditional distribution learning of future neural activity with closed-loop neurotechnology applications."
triggers:
- neural dynamics prediction
- flow matching
- probabilistic forecasting
- brain dynamics
- autoregressive flow
- neural activity prediction
- BOLD signal
- closed loop neurotechnology
- multimodal sensory
paper: "2604.11178"
date_created: "2026-04-23"
---
# 自回归流匹配神经动力学预测方法论 (AFM)
## 概述
自回归流匹配(Autoregressive Flow Matching, AFM)是一种用于神经动力学概率预测的框架。该方法从多模态感官输入学习未来神经活动的条件分布,通过自回归因式分解和连续归一化流实现高质量的概率预测。研究发现,过去BOLD动力学信号是未来神经活动的最主要驱动因素。
## 核心架构
### 1. 问题形式化
```
目标:学习条件分布 p(Y_{future} | X_{past})
其中:
- Y_{future} = [y_{t+1}, y_{t+2}, ..., y_{t+T}] : 未来神经活动
- X_{past} = [x_1, x_2, ..., x_t] : 多模态感官输入历史
- 感官模态包括:视觉、听觉、体感等
概率预测目标:
max Σ log p(y_{t+k} | y_{t+1:t+k-1}, X_{past})
```
### 2. 自回归流匹配(AFM)框架
```
架构组成:
1. 条件编码器 h(X_{past}) → c (上下文向量)
2. 自回归核心:逐步预测未来时间步
3. 流匹配层:每个时间步的条件分布建模
Flow Matching 公式:
给定条件 c 和过去预测 ŷ_{t+1:t+k-1}:
z_0 ~ N(0, I) (噪声样本)
z_1 = y_{t+k} (目标神经活动)
学习向量场 v_θ(z_t, t, c, ŷ_{past}) 使得:
dz_t/dt = v_θ(z_t, t, c, ŷ_{past})
生成过程:z_0 →[v_θ]→ z_1 = ŷ_{t+k}
```
### 3. 条件分布学习
```python
class AutoregressiveFlowMatching(nn.Module):
def __init__(self, dim_neural, dim_sensory, hidden_dim=256, n_flows=4):
super().__init__()
# 多模态感官编码器
self.sensory_encoder = nn.LSTM(
input_size=dim_sensory,
hidden_size=hidden_dim,
num_layers=2,
bidirectional=True
)
# 自回归状态更新器
self.ar_updater = nn.GRUCell(
input_size=dim_neural,
hidden_size=hidden_dim
)
# 流匹配网络
self.flow_nets = nn.ModuleList([
ConditionalFlowNet(hidden_dim, dim_neural)
for _ in range(n_flows)
])
def forward(self, sensory_input, n_future_steps, n_samples=100):
# 编码感官历史
context, _ = self.sensory_encoder(sensory_input)
c = context[-1]
predictions = []
prev_neural = torch.zeros(context.size(1), dim_neural)
for k in range(n_future_steps):
h = self.ar_updater(prev_neural, c)
z_0 = torch.randn(n_samples, *prev_neural.shape)
z_1 = self.flow_match(z_0, h, n_steps=10)
predictions.append(z_1)
prev_neural = z_1.mean(dim=0)
c = h
return torch.stack(predictions)
```
### 4. 关键发现:BOLD动力学的主导驱动作用
```
实验发现:
- 过去BOLD信号解释了未来神经活动变异的 >80%
- 视觉/听觉等特定感官模态的贡献相对较小
- 自回归结构(依赖过去预测)是关键
- BOLD的血流动力学响应函数(HRF)本身提供时间平滑性
启示:
1. 神经活动的自相关性是最强的预测信号
2. 多模态感官输入主要提供"扰动"信息
3. 简单的AR基线已经很强,AFM的优势在于不确定性量化
```
### 5. 闭环神经技术应用
```
AFM在闭环系统中的应用:
1. 实时预测:
- 从当前/过去神经信号预测未来状态
- 提供概率分布而非点估计
- 置信区间用于决策阈值设定
2. 自适应刺激:
- 基于预测的异常检测触发干预
- 预测性刺激时序优化
- 不确定性引导的保守/激进策略切换
3. 临床应用场景:
- 癫痫预测与预防性刺激
- 帕金森病深部脑刺激优化
- 抑郁症闭环神经调控
- 中风康复脑机接口
系统架构:
[神经信号采集] → [AFM预测器] → [概率决策器] → [刺激控制器]
↑ │
└────────────── [反馈信号] ←──────────────────────────┘
```
## 训练流程
```python
def train_afm(model, dataloader, optimizer, n_epochs=100):
for epoch in range(n_epochs):
for batch in dataloader:
sensory_past = batch['sensory']
neural_future = batch['neural_fut']
context = model.encode_sensory(sensory_past)
total_loss = 0
prev_neural = torch.zeros_like(neural_future[0])
for k in range(neural_future.size(0)):
target = neural_future[k]
t = torch.rand(target.size(0), 1)
z_0 = torch.randn_like(target)
z_t = (1 - t) * z_0 + t * target
v_pred = model.predict_velocity(z_t, t, context, prev_neural)
v_target = target - z_0
loss = F.mse_loss(v_pred, v_target)
total_loss += loss
prev_neural = target.detach()
optimizer.zero_grad()
total_loss.backward()
optimizer.step()
```
### 不确定性量化
```python
def predict_with_uncertainty(model, sensory_input, n_future=50, n_samples=500):
predictions = model(sensory_input, n_future, n_samples)
mean = predictions.mean(dim=1)
std = predictions.std(dim=1)
ci_95 = torch.quantile(predictions, torch.tensor([0.025, 0.975]), dim=1)
return {
'mean': mean,
'std': std,
'ci_95_lower': ci_95[0],
'ci_95_upper': ci_95[1],
'samples': predictions
}
```
## 关键优势
1. **概率预测**:提供完整的未来神经活动分布,而非点估计
2. **自回归结构**:自然处理时序依赖性
3. **多模态融合**:整合不同感官模态的信息
4. **不确定性量化**:支持风险感知的闭环决策
5. **灵活的分布建模**:Flow Matching避免分布假设限制
## 应用场景
- **闭环脑机接口**:预测性神经调控
- **癫痫预测**:发作前概率预警
- **神经康复**:个性化恢复轨迹预测
- **药物试验**:预测神经药理效应
- **认知神经科学**:多模态感知-神经映射
## 注意事项
- Flow Matching训练需要足够样本以稳定向量场学习
- 自回归误差累积:长程预测不确定性随时间增长
- BOLD信号的时间分辨率限制(~2s TR)影响实时性
- 闭环应用中预测延迟需要纳入考虑
- 多被试泛化需要个体化微调
## 参考文献
- Paper: 2604.11178 "Probabilistic Prediction of Neural Dynamics via Autoregressive Flow Matching"
- 相关:Flow Matching, Continuous Normalizing Flows, Neural Dynamics Forecasting
## Activation Keywords
- autoregressive-flow-matching-neural
- autoregressive flow matching
- autoregressive flow matching neural
## Tools Used
- `read` - 读取技能文档
- `write` - 创建输出
- `exec` - 执行相关命令
## Instructions for Agents
1. 理解技能的核心方法论
2. 根据用户问题提供针对性回答
3. 遵循最佳实践
## Examples
### Example 1: 基本查询
**User:** 请解释 Autoregressive Flow Matching Neural
**Agent:** Autoregressive Flow Matching Neural 是关于...