Menu
Orthogonal Regularization 方法论用于混合符号-神经动力系统，防止符号结构被神经网络残余吸收，实现互补分解
The review-first prompt is selected by default. You can switch to a direct command or download a local copy.
Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
npx skills add https://github.com/hiyenwong/ai_collection --skill orthoreg-hybrid-symbolic-neural-dynamicsThe command stays on one line. Scroll horizontally to inspect it before copying.
Prefer a local copy? Download the files currently available to SkillsMP.
Based on SOC occupation classification
Showing SKILL.md
| name | orthoreg-hybrid-symbolic-neural-dynamics |
| description | Orthogonal Regularization 方法论用于混合符号-神经动力系统，防止符号结构被神经网络残余吸收，实现互补分解 |
| trigger_words | ["hybrid modeling","symbolic neural","dynamical systems","orthogonal regularization","physics-informed","symbolic discovery","L2 regularization","projection argument","sparse discovery","mechanistic model"] |
| paper_id | arXiv:2606.19145 |
| published | 2026-06-17T00:00:00.000Z |
| authors | Till Richter, Niki Kilbertus |
动力系统建模的持久困境：
混合建模的目标: 结合两者的优势，但面临关键挑战：神经网络可能重新学习符号部分。
标准 L² 正则化基于投影论证：
最小化 ||f_symbolic + f_neural||²
→ 约束 ||f_neural||²
当符号结构通过稀疏发现从数据中学习时：
失败案例:
# 传统方法
f_total = f_symbolic(discovered) + f_neural
# f_neural 可能学到 f_symbolic 已表达的部分
→ 冗余！不可解释！
# Orthogonal Regularization
loss = reconstruction_loss + λ * overlap_penalty
# overlap_penalty = ||⟨f_symbolic, f_neural⟩||²
# 直接测量两个组件的内积
OrthoReg 确保：
设符号组件为 $\phi_s(x)$，神经组件为 $\phi_n(x)$：
$$ \text{OrthoReg} = \lambda \int \phi_s(x) \cdot \phi_n(x) dx $$
$$ \min_{\theta} | \hat{y} - y |^2 + \lambda \cdot \text{Overlap}(\phi_s, \phi_n) $$
其中 Overlap 衡量符号与神经组件的重叠程度。
| 方法 | 适用场景 | 稀疏发现时表现 |
|---|---|---|
| L² 正则化 | 符号结构已知 | 失效（投影论证破坏） |
| OrthoReg | 符号结构学习 | 有效（直接惩罚重叠） |
L² 正则化:
OrthoReg:
在部分库不匹配场景下：
| 指标 | L² 正则化 | OrthoReg |
|---|---|---|
| 符号恢复 | 低 | 高 |
| OOD 行为 | 不稳定 | 稳定 |
| 可解释性 | 低 | 高 |
OrthoReg 在：
import torch
def orthoreg_loss(symbolic_output, neural_output):
"""
计算符号与神经输出的正交性损失
Args:
symbolic_output: 符号模型输出
neural_output: 神经网络残余输出
Returns:
overlap_penalty: 重叠惩罚
"""
# 内积度量重叠
overlap = torch.sum(symbolic_output * neural_output)
# L² 形式的重叠惩罚
overlap_penalty = overlap ** 2
return overlap_penalty
class HybridDynamicalModel:
def forward(self, x):
# 符号组件（稀疏发现）
symbolic = self.symbolic_library(x)
# 神经残余
residual = self.neural_net(x)
# 组合输出
return symbolic + residual
def loss(self, pred, target, lambda_ortho=0.1):
# 重构损失
recon = F.mse_loss(pred, target)
# 正交约束
symbolic = self.symbolic_library(self.x)
residual = self.neural_net(self.x)
ortho = orthoreg_loss(symbolic, residual)
return recon + lambda_ortho * ortho
$$ f_{total} = f_s + f_n $$
其中：
| 方法 | 符号发现支持 | 重叠预防 | 可解释性保证 |
|---|---|---|---|
| SINDy | ✓ | 间接 | ✓ |
| L² 正则化 | ✗（投影失效） | ✗ | 低 |
| OrthoReg | ✓ | ✓ | ✓ |
Till Richter, Niki Kilbertus. "OrthoReg: Orthogonal Regularization for Hybrid Symbolic-Neural Dynamical Systems." arXiv:2606.19145. 2026-06-17.