Files
myWiki/concepts/kl-order.md

1.4 KiB
Raw Blame History

title, created, updated, type, tags, sources
title created updated type tags sources
KL 阶 (KL Order) 2026-06-10 2026-06-10 concept
singular-learning-theory
information-geometry
bridge-invariant
dead-directions-geometric-singular-learning

KL 阶 (KL Order)

KL 阶 k 是沿 dead-direction 接近奇异集时 KL 散度趋于零的速率:

K(theta(t)) = c · t^{2k} + O(t^{2k+1})

KL 散度在 t 处有 2k 阶零点——两倍于 KL 阶。

桥接不变量

KL 阶是singular-learning-theoryinformation-geometry都可计算的少数不变量之一:

与 Deep Direction Fisher 衰减的关系

k = 1: Fisher decay rate 0 (正则方向)
k = 2: Fisher decay rate 2
k = 3: Fisher decay rate 4

从 Checkpoint 计算

Shirodkar (2026) 证明可通过一次前向+反向传播计算 KL 阶——无需后验采样,无需消解。这使 SLT 分析在大规模网络上首次进入实践领域。

参考