Files
myWiki/concepts/natural-gradient-descent.md

34 lines
1.3 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

---
title: "自然梯度下降"
created: 2026-06-22
updated: 2026-06-22
type: concept
tags: [optimization, information-geometry, manifold-learning]
sources: [nano-filter]
---
# 自然梯度下降
Natural gradient descent 是考虑参数空间几何结构的梯度下降方法。不同于标准梯度下降在欧氏空间中取最陡下降方向,自然梯度利用 **Fisher 信息矩阵** $F$ 调整梯度方向以适配参数空间的曲率Riemannian 结构)。
## 更新公式
$$
v^{(i+1)} = v^{(i)} - \eta F_v^{-1} \frac{\partial J}{\partial v}\Big|_{v=v^{(i)}}
$$
其中 $F_v$ 是 Fisher 信息矩阵。与标准梯度下降 $v - \eta \nabla J$ 的区别在于用 $F_v^{-1}$ 对梯度做度量校正。
## 在高斯流形上的应用
[[nano-filter|NANO filter]] 的核心创新:在 [[gaussian-manifold|高斯流形]] 上执行自然梯度下降,直接最小化更新步的优化目标 $J(\hat{x}_t, P_t)$,避免传统 Gaussian filter 的线性化误差。Fisher 矩阵在高斯分布 $N(x; \hat{x}_t, P_t)$ 下具有解析形式:
$$
F_v^{-1} = \begin{bmatrix} P_t & 0 \\ 0 & 2(P_t^{-1} \otimes P_t^{-1}) \end{bmatrix}
$$
## 参考
- [[gaussian-manifold|Gaussian Manifold]]
- [[fisher-information-metric|Fisher Information Metric]]
- [[nano-filter|NANO Filter]]