Problems 专栏收录该内容
3 篇文章 0 订阅

# RBF 函数定义

κ ( x 1 , x 2 ) = < ϕ ( x 1 ) , ϕ ( x 2 ) > = exp ⁡ ( − ∣ ∣ x 1 − x 2 ∣ ∣ 2 2 σ 2 ) \kappa(\mathbf{x_1}, \mathbf{x_2}) = \left<\phi(\mathbf{x_1}), \phi(\mathbf{x_2})\right>=\exp\left(-\frac{||\mathbf{x_1} - \mathbf{x_2}||^2}{2\sigma^2}\right)

# RBF 函数性质

1. 所有的 ϕ ( x ) \phi(\mathbf{x}) 都是 x \mathbf{x} 被映射到无穷维空间后的一个在超球面上的点，因为
∣ ∣ ϕ ( x ) − 0 ∣ ∣ 2 = ∣ ∣ ϕ ( x ) ∣ ∣ 2 = < ϕ ( x ) , ϕ ( x ) > = exp ⁡ ( − ∣ ∣ x − x ∣ ∣ 2 2 σ 2 ) = exp ⁡ ( 0 ) = 1 ||\phi(\mathbf{x}) - \mathbf{0}||^2 = ||\phi(\mathbf{x}) ||^2 = \left<\phi(\mathbf{x}), \phi(\mathbf{x})\right> = \exp\left(-\frac{||\mathbf{x} - \mathbf{x}||^2}{2\sigma^2}\right) = \exp(0) = 1
2. 无穷维空间导致计算运行开销耗时较大，训练时间长，虽然kernel函数已经将计算量核存储空间减小了

# 为什么 RBF 函数是将 x \mathbf{x} 被映射到无穷维后构造的核函数？

< ϕ ( x 1 ) , ϕ ( x 2 ) > = exp ⁡ ( − ∣ ∣ x 1 − x 2 ∣ ∣ 2 2 σ 2 ) = exp ⁡ ( − ∣ ∣ x 1 ∣ ∣ 2 2 σ 2 ) exp ⁡ ( − ∣ ∣ x 2 ∣ ∣ 2 2 σ 2 ) exp ⁡ ( x 1 ⋅ x 2 σ 2 ) \left<\phi(\mathbf{x_1}), \phi(\mathbf{x_2})\right> = \exp\left(-\frac{||\mathbf{x_1} - \mathbf{x_2}||^2}{2\sigma^2}\right) \\ = \exp(-\frac{||\mathbf{x_1}||^2}{2\sigma^2})\exp(-\frac{||\mathbf{x_2}||^2}{2\sigma^2})\exp(\frac{\mathbf{x_1\cdot x_2}}{\sigma^2})\\

exp ⁡ ( x 1 ⋅ x 2 σ 2 ) = Σ i = 1 ∞ ( x 1 ⋅ x 2 σ 2 ) i i ! \exp(\frac{\mathbf{x_1\cdot x_2}}{\sigma^2}) = \Sigma_{i=1}^{\infty}\frac{(\frac{\mathbf{x_1\cdot x_2}}{\sigma^2})^i}{i!}

< ϕ ( x 1 ) , ϕ ( x 2 ) > = Σ i = 0 ∞ ( x 1 T ⋅ x 2 σ 2 ) i i ! exp ⁡ ( − ∣ ∣ x 1 ∣ ∣ 2 2 σ 2 ) exp ⁡ ( − ∣ ∣ x 2 ∣ ∣ 2 2 σ 2 ) = Σ i = 0 ∞ ( x 1 T σ ) i i ! exp ⁡ ( − ∣ ∣ x 1 ∣ ∣ 2 2 σ 2 ) ⋅ ( x 2 σ ) i i ! exp ⁡ ( − ∣ ∣ x 2 ∣ ∣ 2 2 σ 2 ) = Σ i = 0 ∞ ( ( x 1 σ 2 ) i i ! exp ⁡ ( − ∣ ∣ x 1 ∣ ∣ 2 2 σ ) ) T ⋅ ( x 2 σ ) i i ! exp ⁡ ( − ∣ ∣ x 2 ∣ ∣ 2 2 σ 2 ) = ϕ T ( x 1 ) ⋅ ϕ ( x 2 ) \left<\phi(\mathbf{x_1}), \phi(\mathbf{x_2})\right> = \Sigma_{i=0}^{\infty}\frac{(\frac{\mathbf{x_1^T\cdot x_2}}{\sigma^2})^i}{i!}\exp(-\frac{||\mathbf{x_1}||^2}{2\sigma^2})\exp(-\frac{||\mathbf{x_2}||^2}{2\sigma^2})\\ =\Sigma_{i=0}^{\infty} \frac{(\frac{\mathbf{x_1^T}}{\sigma})^i}{\sqrt{i!}}\exp(-\frac{||\mathbf{x_1}||^2}{2\sigma^2})\cdot \frac{(\frac{\mathbf{x_2}}{\sigma})^i}{\sqrt{i!}}\exp(-\frac{||\mathbf{x_2}||^2}{2\sigma^2})\\ =\Sigma_{i=0}^{\infty}(\frac{(\frac{\mathbf{x_1}}{\sigma^2})^i}{\sqrt{i!}}\exp(-\frac{||\mathbf{x_1}||^2}{2\sigma}))^T \cdot\frac{(\frac{\mathbf{x_2}}{\sigma})^i}{\sqrt{i!}}\exp(-\frac{||\mathbf{x_2}||^2}{2\sigma^2})\\=\phi^T(\mathbf{x_1})\cdot \phi(\mathbf{x_2})

ϕ i ( x ) = ( x i i ! σ i ) exp ⁡ ( − ∣ ∣ x ∣ ∣ 2 2 σ 2 ) \phi_i(\mathbf{x}) = (\frac{\mathbf{x^i}}{\sqrt{i!}\sigma^i})\exp(-\frac{||\mathbf{x}||^2}{2\sigma^2})
ϕ ( x ) = [ ϕ 0 ( x ) , . . . , ϕ i ( x ) , . . . , ϕ ∞ ( x ) ] \phi(\mathbf{x}) = [\phi_0(\mathbf{x}) , ...,\phi_i(\mathbf{x}) , ..., \phi_{\infty}(\mathbf{x}) ]

• 0
点赞
• 0
评论
• 5
收藏
• 一键三连
• 扫一扫，分享海报

06-05 1586
03-12 527

04-17 3万+
11-21 580
05-20 577
06-26 3122
10-19 150
06-29 5826
11-16 1万+
07-26 6735
©️2021 CSDN 皮肤主题: 1024 设计师:白松林

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、C币套餐、付费专栏及课程。