Files
myWiki/concepts/pseudo-huber-loss.md

32 lines
1.2 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: "Pseudo-Huber 损失"
created: 2026-06-22
updated: 2026-06-22
type: concept
tags: [robust-statistics, loss-function, optimization]
sources: [nano-filter]
---
# Pseudo-Huber 损失
Pseudo-Huber loss 是 Huber loss 的光滑可微变体用于鲁棒回归和离群值处理。在小残差时表现如平方损失L2在大残差时切换为线性增长L1从而抑制离群值对梯度的过度影响。
## 定义
$$
\ell_{ph}(x_t, y_t) = \delta^2 \left( \sqrt{1 + (y_t - g(x_t))^2 / \delta^2} - 1 \right)
$$
其中 $\delta > 0$ 控制从二次到线性增长转变的阈值。
## 在 NANO 鲁棒变体中的作用
[[nano-filter|NANO filter]] 的鲁棒变体(基于 [[gibbs-posterior|Gibbs 后验]]框架)采用 Pseudo-Huber 损失替代负对数似然:
- 当测量模型误设(传感器故障、未建模动态)时,标准平方损失会因离群值产生巨大梯度
- Pseudo-Huber 损失平滑地截断这种影响,保持估计的可靠性
- 与标准 Huber 损失相比Pseudo-Huber 处处可微,更适合基于梯度的优化(如 [[natural-gradient-descent|自然梯度]]
## 参考
- [[gibbs-posterior|Gibbs Posterior]]
- [[nano-filter|NANO Filter]]