1、正则化网络
每个隐藏单元的激活函数由Green函数定义
(式1)G(x,xi)=exp(−12σi2∣∣x−xi∣∣2)G(x,x_i) = exp(- \frac{1}{2\sigma _i ^2}||x-x_i||^2) \tag {式1}G(x,xi)=exp(−2σi21∣∣x−xi∣∣2)(式1)
2、广义径向基函数网络
F∗(x)=∑i=1m1wiφ(x,ti)F^*(x) = \sum _{i=1}^{m_1} w_i \varphi(x,t_i)F∗(x)=i=1∑m1wiφ(x,ti)
φ(x,ti)=G(∣∣x−ti∣∣)\varphi(x,t_i) = G(||x-t_i||)φ(x,ti)=G(∣∣x−ti∣∣)
F∗(x)=∑i=1m1wiφ(x,ti)=∑i=1m1wiG(x,ti)=∑i=1m1wiG(∣∣x−ti∣∣)F^*(x) = \sum _{i=1}^{m_1} w_i \varphi(x,t_i) = \sum _{i=1}^{m_1} w_i G(x,t_i)= \sum _{i=1}^{m_1} w_i G(||x-t_i||)F∗(x)=i=1∑m1wiφ(x,ti)=i=1∑m1wiG(x,ti)=i=1∑m1wiG(∣∣x−ti∣∣)
新的代价函数:
其中:
∣∣DF∗∣∣2=<DF∗,DF∗>H=[∑i=1m1wiG(x,ti),D~D∑i=1miwiG(X,ti)]H||DF^*||^2 = <DF^*,DF^*>_H = [\sum _{i=1} ^ {m_1} w_iG(x,t_i),\tilde{D}D\sum _{i=1}^{m_i}w_i G(X,t_i)]_H∣∣DF∗∣∣2=<DF∗,DF∗>H=[i=1∑m1wiG(x,ti),D~Di=1∑miwiG(X,ti)]H
=[∑i=1m1wiG(X,ti),∑i=1m1w)δti]H=∑j=1m1∑i=1m1wjwiG(tj,ti)=WTG0W=[\sum _{i=1} ^{m_1} w_iG(X,t_i),\sum _{i=1}^{m_1} w )\delta_{t_i} ]_H = \sum_{j=1}^{m_1} \sum_{i=1}^{m_1}w_j w_iG(t_j,t_i)=W^TG_0W=[i=1∑m1wiG(X,ti),i=1∑m1w)δti]H=j=1∑m1i=1∑m1wjwiG(tj,ti)=WTG0W
3、加权范数
∣∣X∣∣C2=(CX)T(CX)=XTCTCX||X|| _C ^2 = (CX)^T(CX) = X^TC^TCX∣∣X∣∣C2=(CX)T(CX)=XTCTCX
F∗(x)=∑i=1m1wiG(∣∣x−ti∣∣c)F^*(x) = \sum _{i=1}^{m_1} w_i G(||x-t_i||_c) F∗(x)=i=1∑m1wiG(∣∣x−ti∣∣c)
一个以tit_iti为中心和具有范数加权矩阵C的高斯径向基函数G(∣∣X−ti∣∣c)G(||X-t_i||_c)G(∣∣X−ti∣∣c)可写成
G(∣∣X−ti∣∣c=exp(−(X−ti)TCTC(X−ti)]=exp[−12(X−ti)TΣ−1(X−ti)]))G(||X-t_i||_c = exp(-(X-t_i)^TC^TC(X-t_i)] = exp[-\frac{1}{2}(X-t_i)^T \Sigma ^{-1} (X-t_i) ]))G(∣∣X−ti∣∣c=exp(−(X−ti)TCTC(X−ti)]=exp[−21(X−ti)TΣ−1(X−ti)]))
12Σ−1=CTC\frac{1}{2} \Sigma ^{-1} = C^TC21Σ−1=CTC