Skip to content

Latest commit

 

History

History
51 lines (34 loc) · 1.06 KB

Week2-MSE.md

File metadata and controls

51 lines (34 loc) · 1.06 KB

Week 2: Mathematical Symbols Explanations

Basic Notations

Set $z$.

$P(x|z)$ means get an element $x$ from set $z$ and apply function $P$ on it.

GMM

Formulas

$$P(x)$$ $x$: lower case, one sample

$$ \begin{aligned} P(X) &= P(x_1)\cdot P(x_2)\cdots P(x_N) \&= \prod^N_{i=1}P(x_i) \&= \prod_nP(x_n) \end{aligned} $$ $X$: upper case, set of samples, i.e. $x_1, x_2, \cdots, x_N$

Parameters

$\mu$: 决定了分布的中心点在哪

$\Sigma$: 标准差,决定了函数图像的形状,i.e. 是更圆滑还是更尖锐

$\pi$: 每个集群的权重,i.e.: 在 $k$ 个集合中, $\pi_k=P(Z_k)$

GMM 是为了寻找 $\mu, \Sigma, \pi$ 的最佳组合。

Transformation

考虑到最终 $P(X)$ 公式中 $\prod_nP(x_n)$ 是难以处理的,因此我们需要进行优化:

$$ \begin{aligned} &\ln P(X) = \ln \left(\prod_nP(x_n)\right) \\ =& \sum_n \ln P(x_n) \\ =& \sum_n \ln \sum_z (P(x|z) \pi_z)\\ =& \sum_n \ln \left( \sum_k\left(\pi_k \mathcal{N}(x_n|\mu_k, \Sigma_k)\right) \right) \end{aligned} $$

延申自 $\ln (ab)=\ln a + \ln b$