最近在看深度学习的文章,发现它们经常会提到Lipschitz条件,特此做个记录。
Lipschitz continuous
Link
- In particular, a real-valued function f : R → R is called Lipschitz continuous if there exists a positive real constant K such that, for all real x1 and x2,
∣
f
(
x
1
)
?
f
(
x
2
)
∣
≤
K
∣
x
1
?
x
2
∣
,
{\displaystyle |f(x_{1})-f(x_{2})|\leq K|x_{1}-x_{2}|,}
∣f(x1?)?f(x2?)∣≤K∣x1??x2?∣,where K is called Lipschitz constant.
- 可以证明,导函数有界的函数都满足Lipschitz连续。
Lipschitz约束的一个应用
Link
Reference: Spectral Norm Regularization for Improving the Generalizability of Deep Learning
- L2 正则化能使得模型更好地满足Lipschitz条件,从而降低模型对输入扰动的敏感性,增强模型的泛化性能。但这只是一个比较粗糙的条件,更精确的条件应该是谱范数(Spectral Norm)。
- Spectral Norm
∣
∣
W
∣
∣
2
||W||_2
∣∣W∣∣2?:
W
T
W
W^{T}W
WTW的最大特征根的平方根。(
W
W
W是神经网络的权重矩阵)
- Spectral Norm Regularization: 把谱范数的平方作为额外的正则项,取代简单的L2 正则项。
- 注:柯西不等式、幂迭代(power iteration)
|