机器学习白板推导（2）

机器学习白板推导（1）

八、概率图模型

4.Bayesian Network-具体模型

$$BayesianNetwork\{\begin{array}{l}单一：NaiveNayes⟶P(x\mid y)={\prod }_{i=1}^{p}p(x_i\mid y=1)\\ \begin{array}{l}混合：GMM\\ 时间：\{\begin{array}{l}MarkovChain\\ GuassianProcess(无限维高斯分布)\end{array}\end{array}\}动态模型\\ 连续：GuassianBayesianNetwork\end{array}$$

$$动态模型\{\begin{array}{l}HMM(离散状态)\\ LDS(KalmanFilter,)(连续，线性)\\ ParitceFilter\end{array}$$

Naive Nayes
tail to tail

GMM
$z$是离散的

5.马尔可夫随机场-Representation-条件独立性

有向图的局部结构head to head ，默认是独立的，观测反而不独立

马尔可夫随机场是无向图，则没有head to head 的困扰

条件独立性体现在三个方面：

• Global Markov全局马尔可夫
  $X_A\perp X_C\mid X_B$

• Local Markov局部马尔可夫
  a ⊥ { e , f } ∣ { b , c , d } a\bot \left \{ e,f \right \} \mid \left \{ b,c,d \right \} a⊥{ e,f}∣{ b,c,d}

• 成对Markov
  $x_i\perp x_j\mid x_{-i},x_{-j}(i\ne j,i,j)$不相邻

三个是等价的

引入团的概念

团：一个关于节点的集合，集合的节点之间都是连通（两两）

$$\begin{array}{rl}P(x)& =\frac{1}{Z}\prod _{i=1}^K\phi (x_{C_i})\\ Z& =\sum _x\prod _{i=1}^K\phi (x_{C_i})\\ & =\sum _{x_1}\sum _{x_2}\cdots \sum _{x_p}\prod _{i=1}^K\phi (x_{C_i})\end{array}$$
$C_i$: 最大团
$x_{C_i}$: 最大团随机变量集合
$\phi (x_{C_i})$: 势函数，必须为正

6.马尔可夫随机场-Representation-因子分解

Hammesley-Clifford 定理可以证明因子分解（基于最大团）与条件独立性等价
φ ( x C i ) = exp ⁡ { − E ( x C i ) } > 0 \varphi (x_{C_i})=\exp \left \{ -E(x_{C_i}) \right \} >0 φ(xCi​​)=exp{ −E(xCi​​)}>0$E(x_{C_i})$为能量函数
$P(x)$称为Gibbs分布
P ( x ) = 1 Z ∏ i = 1 K φ ( x C i ) = 1 Z ∏ i = 1 K exp ⁡ { − E ( x C i ) } = 1 Z exp ⁡ { − ∑ i = 1 K E ( x C i ) } \begin{aligned} P(x) &=\frac{1}{Z} \prod_{i=1}^{K} \varphi (x_{C_i}) \\ &=\frac{1}{Z} \prod_{i=1}^{K} \exp \left \{ -E(x_{C_i}) \right \}\\ &=\frac{1}{Z} \exp \left \{ -\sum_{i=1}^K E(x_{C_i}) \right \} \end{aligned} P(x)​=Z1​i=1∏K​φ(xCi​​)=Z1​i=1∏K​exp{ −E(xCi​​)}=Z1​exp{ −i=1∑K​E(xCi​​)}​指数族分布

最大熵原理$\Rightarrow$指数族分布(Gibbs分布)

Markov Random Field $\iff$ Gibbs Distribution

7.推断-总体介绍

$P(x)=P(x_1,x_2,\cdots ,x_p)$

推断-求概率

• 边缘概率$$P(x_i)=\sum _{x_1}\cdots \sum _{x_{i-1}}\sum _{x_{i+1}}\cdots \sum _{x_p}P(x)$$
• 条件概率$$P(x_A\mid x_B),x=x_A\cup x_B$$
• MAP Inference:$$\hat{z}=\arg \underset{z}{\max }P(z\mid x)\propto \arg \underset{z}{\max }P(x,z)$$

$$\{\begin{array}{l}精确推断\{\begin{array}{l}VariableElimanation(VE)\\ BeliefPropagtion(BP)⟶SumProductAlgorithm\\ JunctionTreeAlgorithm\end{array}\\ 近似推断\{\begin{array}{l}LoopBeliefPropagation\\ MenteCarloInference：ImportanceSampling,MCMC\\ VariationalInference\end{array}\end{array}$$

8.推断-Variable Elimination

求边缘概率
假设a,b,c,d均是离散的二值r.v，a,b,c,d ∈ { 0 , 1 } \in \left \{ 0,1 \right \} ∈{ 0,1}

$$\begin{gathered} P(d)=\sum _{a,b,c}P(a)P(b)P(c)P(d) \\ =\sum _{a,b,c}P(a)P(b\mid a)P(c\mid b)P(d\mid c) \end{gathered}$$
逐一带入计算，P(d)计算量为$2^3$，，随机变量取$K$种$p$维，计算量为幂指函数$K^p$，显然不可取。

$$\begin{array}{rl}P(d)& =\sum _{a,b,c}P(a)P(b\mid a)P(c\mid b)P(d\mid c)\\ & =\sum _{b,c}P(c\mid b)P(d\mid c)\sum _{a}P(a)P(b\mid a)\\ & =\sum _{c}P(d\mid c)P(c\mid b)\sum _b{\varphi }_a(b)\\ & =\sum _{c}P(d\mid c){\varphi }_b(c)\\ & ={\varphi }_c(d)\\ P(x)& =\prod _{x_c}{\phi }_c(x_c)\end{array}$$乘法分配率

缺点：重复计算（下节的内容），消除的次序

最大团：不可再添加节点，使得团内节点兩两两连接，团之间的联系非常小，可以认为是相互独立的

9.推断-Variable Elimination to Blief Propagation

$$P(e)=\sum _{d}p(e\mid d)\sum _{c}P(d\mid c)\sum _{b}P(c\mid b)\sum _{a}P(b\mid a)P(a)$$
同理
$$P(c)=\left(\sum _{b}P(c\mid b)\sum _{a}P(b\mid a)P(a)\right)\left(\sum _{d}p(d\mid c)\sum _{e}P(e\mid d)\right)$$
如何计算万$P(e)$，又要计算$P(c)$，等式有公因子，如果再计算其他任务就会造成重复计算。

---

Chain链式 to Tree树结构（如计算$P(c)$，前向-反向算法）
有向 to 无向

联合概率
$$P(a,b,c,d)=\frac{1}{Z}{\phi }_a(a){\phi }_b(b){\phi }_c(c){\phi }_d(d){\phi }_{a}b(ab){\phi }_{b}c(bc){\phi }_{b}d(bd)$$
求边缘概率
$$P(a)=\sum _{b,c,d}P(a,b,c,d)$$
$$P(a)\{\begin{array}{l}{\phi }_a\\ m_{b\to a}(a)\{\begin{array}{l}{\sum }_b\\ m_{c\to b}(b)\{\begin{array}{l}{\sum }_c{\phi }_c{\phi }_{bc}\end{array}\\ {\phi }_b\\ m_{c\to b}(b)\{\begin{array}{l}{\sum }_d{\phi }_d{\phi }_{bd}\end{array}\\ {\phi }_{ab}\end{array}\end{array}$$

$$\{\begin{array}{l}{m}_{b\to a}(x_a)={\sum }_{x_b}{\phi }_{ab}{\phi }_b{m}_{c\to b}(x_b)m_{c\to b}(x_b)\\ \\ P(x_a)={\phi }_a{m}_{b\to a}(x_a)\end{array}$$
$$\{\begin{array}{l}{m}_{j\to i}(x_i)={\sum }_{x_j}{\phi }_{ij}{\phi }_j{\prod }_{k\in NB(j)-i}{m}_{k\to j}(x_j)\\ \\ P(x_i)={\phi }_i{\prod }_{k\in NB(i)}{m}_{k\to i}(x_i)\end{array}$$
$NB(j)-i,j$的邻节点除了$i$节点

不要直接求边缘概率，只需求$m_{i\to j}$

10.推断-Blief Propagation

$$\{\begin{array}{l}beilef(b)={\phi }_{b}child\\ \\ m_{b\to a}={\sum }_b{\phi }_{ab}belief(b)\end{array}$$

BP = VE + Caching
直接求$m_{ij}$
图的遍历

• Get root, assume is root
• Collect Message

for x_i in NB(Root):
	collect(x_i)

• Distribution Message

for x_j  in NB(Root):
	distribute(x_j)

可以并行运算

11.Max Product

$$\{\begin{array}{l}{m}_{j\to i}(x_i)={\sum }_{x_j}{\phi }_{ij}{\phi }_j{\prod }_{k\in NB(j)-i}{m}_{k\to j}(x_j)\\ \\ P(x_i)={\phi }_i{\prod }_{k\in NB(i)}{m}_{k\to i}(x_i)\end{array}$$

将Sum Product $\sum$符改为Max Product $\max$

• BP的改进
• Viterbi的推广

路径最优，每层的每一个节点都有累积最短路径
m j → i = max ⁡ x j φ j ⋅ φ i j ∏ k ∈ N B ( j ) − 1 m k → j m c → b = max ⁡ x c φ c ⋅ φ b c m d → b = max ⁡ x d φ d ⋅ φ b d m b → a = max ⁡ x b φ b ⋅ φ a b m c → b m d → b \begin{aligned} m_{j \to i}&=\max_{x_j} \varphi_j \cdot \varphi_{ij} \prod_{k \in NB(j)-1} m_{k \to j}\\ m_{c \to b}&=\max_{x_c} \varphi_c \cdot \varphi_{bc}\\ m_{d \to b}&=\max_{x_d} \varphi_d \cdot \varphi_{bd}\\ m_{b \to a}&=\max_{x_b} \varphi_b \cdot \varphi_{ab} m_{c \to b} m_{d \to b} \end{aligned} mj→i​mc→b​md→b​mb→a​​=xj​max​φj​⋅φij​k∈NB(j)−1∏​m