机器学习笔记之高斯混合模型——EM算法求解高斯混合模型【E步操作】
引言
上一节介绍了尝试使用极大似然估计求解高斯混合模型的模型参数，但无法求出解析解。本节将介绍使用EM算法求解高斯混合模型的模型参数。
回顾：高斯混合模型及模型参数
令
X
\mathcal X
X表示观测数据(Observed Data)，共包含
N
N
N个样本点，并假设 任意样本之间独立同分布：
X
=
{
x
(
1
)
,
x
(
2
)
,
⋯
,
x
(
N
)
}
x
(
i
)
∼
i.i.d.
x
(
j
)
(
x
(
i
)
,
x
(
j
)
∈
X
;
i
≠
j
)
\mathcal X = \{x^{(1)},x^{(2)},\cdots, x^{(N)}\} \\ x^{(i)} \overset{\text{i.i.d.}}{\sim}x^{(j)} \quad (x^{(i)},x^{(j)} \in \mathcal X;i\neq j)
X={x(1),x(2),⋯,x(N)}x(i)∼i.i.d.x(j)(x(i),x(j)∈X;i=j)
任意一个样本点
x
(
i
)
x^{(i)}
x(i)均对应一个隐变量
z
(
i
)
z^{(i)}
z(i)。从样本数量角度观察，隐变量集合
Z
\mathcal Z
Z表示如下：
Z
=
{
z
(
1
)
,
z
(
2
)
,
⋯
,
z
(
N
)
}
\mathcal Z = \{z^{(1)},z^{(2)},\cdots,z^{(N)}\}
Z={z(1),z(2),⋯,z(N)}
称
(
X
,
Z
)
(\mathcal X,\mathcal Z)
(X,Z)为完整数据(Complete Data)，样本数量角度表示如下：
(
X
,
Z
)
=
{
(
x
(
1
)
,
z
(
1
)
)
,
⋯
,
(
x
(
N
)
,
z
(
N
)
)
}
(\mathcal X,\mathcal Z) = \{(x^{(1)},z^{(1)}),\cdots,(x^{(N)},z^{(N)})\}
(X,Z)={(x(1),z(1)),⋯,(x(N),z(N))}
从变量分布的角度观察，隐变量
Z
\mathcal Z
Z是基于
K
\mathcal K
K个参数的离散分布，各参数及对应概率分布表示如下：
| Z \mathcal Z Z | z 1 z_1 z1 | z 2 z_2 z2 | ⋯ \cdots ⋯ | z K z_{\mathcal K} zK |
|---|---|---|---|---|
| P ( Z ) P(\mathcal Z) P(Z) | p 1 p_1 p1 | p 2 p_2 p2 | ⋯ \cdots ⋯ | p K p_{\mathcal K} pK |
并满足：
∑
k
=
1
K
p
k
=
1
\sum_{k=1}^{\mathcal K} p_k = 1
k=1∑Kpk=1
任意
z
j
∈
Z
z_j \in \mathcal Z
zj∈Z均唯一对应一个高斯分布。换句话说，给定隐变量标签
z
j
∈
Z
z_j \in \mathcal Z
zj∈Z的条件下，
z
j
z_j
zj标签下的样本数据
x
x
x服从高斯分布。因而共包含
K
\mathcal K
K个高斯分布：
| Z \mathcal Z Z | z 1 z_1 z1 | z 2 z_2 z2 | ⋯ \cdots ⋯ | z K z_{\mathcal K} zK |
|---|---|---|---|---|
| P ( X ∣ Z ) P(\mathcal X \mid \mathcal Z) P(X∣Z) | N ( μ 1 , Σ 1 ) \mathcal N(\mu_1,\Sigma_1) N(μ1,Σ1) | N ( μ 2 , Σ 2 ) \mathcal N(\mu_2,\Sigma_2) N(μ2,Σ2) | ⋯ \cdots ⋯ | N ( μ K , Σ K ) \mathcal N(\mu_{\mathcal K},\Sigma_{\mathcal K}) N(μK,ΣK) |
数学符号表达即：
P
(
X
∣
Z
=
z
k
)
∼
N
(
X
∣
μ
k
,
Σ
k
)
(
k
=
1
,
2
,
⋯
,
K
)
P(\mathcal X\mid \mathcal Z = z_k) \sim \mathcal N(\mathcal X \mid \mu_{k},\Sigma_k) \quad (k=1,2,\cdots,\mathcal K)
P(X∣Z=zk)∼N(X∣μk,Σk)(k=1,2,⋯,K)
因此，高斯混合模型的概率模型
P
(
X
)
P(\mathcal X)
P(X)表达如下：
P
(
X
)
=
∑
Z
P
(
X
∣
Z
=
z
k
)
P
(
Z
=
z
k
)
=
∑
k
=
1
K
N
(
X
∣
μ
k
,
Σ
k
)
⋅
p
k
\begin{aligned}P(\mathcal X) & = \sum_{\mathcal Z} P(\mathcal X \mid \mathcal Z = z_k)P(\mathcal Z = z_k) \\ & = \sum_{k=1}^{\mathcal K}\mathcal N(\mathcal X \mid \mu_k,\Sigma_k)\cdot p_k \end{aligned}
P(X)=Z∑P(X∣Z=zk)P(Z=zk)=k=1∑KN(X∣μk,Σk)⋅pk
概率模型的模型参数
θ
\theta
θ表示如下：
θ
=
{
p
1
,
⋯
,
p
K
,
μ
1
,
⋯
,
μ
K
,
Σ
1
,
⋯
,
Σ
K
}
\theta = \{p_1,\cdots,p_{\mathcal K},\mu_1,\cdots,\mu_{\mathcal K},\Sigma_1,\cdots,\Sigma_{\mathcal K}\}
θ={p1,⋯,pK,μ1,⋯,μK,Σ1,⋯,ΣK}
回顾：狭义EM算法
EM算法是求解概率模型
P
(
X
∣
θ
)
P(\mathcal X \mid \theta)
P(X∣θ)模型参数的一种方法，它的底层是极大似然估计，它的迭代求解公式具体表示如下：
θ
(
t
+
1
)
=
arg
max
θ
∫
Z
log
P
(
X
,
Z
∣
θ
)
⋅
P
(
Z
∣
X
,
θ
(
t
)
)
d
Z
=
arg
max
θ
E
Z
∣
X
,
θ
[
log
P
(
X
,
Z
∣
θ
)
]
\begin{aligned} \theta^{(t+1)} & = \mathop{\arg\max}\limits_{\theta} \int_{\mathcal Z} \log P(\mathcal X,\mathcal Z \mid \theta) \cdot P(\mathcal Z \mid \mathcal X,\theta^{(t)}) d\mathcal Z \\ & = \mathop{\arg\max}\limits_{\theta} \mathbb E_{\mathcal Z \mid \mathcal X,\theta} \left[\log P(\mathcal X,\mathcal Z \mid \theta)\right] \end{aligned}
θ(t+1)=θargmax∫ZlogP(X,Z∣θ)⋅P(Z∣X,θ(t))dZ=θargmaxEZ∣X,θ[logP(X,Z∣θ)]
基于上述公式，可以将EM算法分成两个步骤：
- E步(Expection-Step)：令
E
Z
∣
X
,
θ
[
log
P
(
X
,
Z
∣
θ
)
]
E_{\mathcal Z \mid \mathcal X,\theta} \left[\log P(\mathcal X,\mathcal Z \mid \theta)\right]
EZ∣X,θ[logP(X,Z∣θ)]表示为 关于
θ
,
θ
(
t
)
\theta,\theta^{(t)}
θ,θ(t)的函数。则有：
L ( θ , θ ( t ) ) = ∫ Z log P ( X , Z ∣ θ ) ⋅ P ( Z ∣ X , θ ( t ) ) d Z \mathcal L(\theta,\theta^{(t)}) = \int_{\mathcal Z} \log P(\mathcal X,\mathcal Z \mid \theta) \cdot P(\mathcal Z \mid \mathcal X,\theta^{(t)}) d\mathcal Z L(θ,θ(t))=∫ZlogP(X,Z∣θ)⋅P(Z∣X,θ(t))dZ - M步(Maximization-Step)：基于E步操作，选择合适的
θ
\theta
θ，使得
L
(
θ
,
θ
(
t
)
)
\mathcal L(\theta,\theta^{(t)})
L(θ,θ(t))最大。
θ ( t + 1 ) = arg max θ L ( θ , θ ( t ) ) \theta^{(t+1)} = \mathop{\arg\max}\limits_{\theta} \mathcal L(\theta,\theta^{(t)}) θ(t+1)=θargmaxL(θ,θ(t))
E步、M步交替进行，最终迭代收敛至最优解(至少局部最优)。
使用EM算法求解高斯混合模型参数
场景整理
EM算法中符号表示与高斯混合模型中的符号表示 对比如下：
等号左端是‘EM算法’的符号表示；等号右端是‘高斯混合模型’的符号表示。
P
(
X
,
Z
∣
θ
)
=
P
(
Z
=
z
j
)
⋅
P
(
X
∣
Z
=
z
j
)
=
p
Z
⋅
N
(
X
∣
μ
Z
,
Σ
Z
)
=
∏
i
=
1
N
p
z
(
i
)
⋅
N
(
x
(
i
)
∣
μ
z
(
i
)
,
Σ
z
(
i
)
)
P
(
Z
∣
X
,
θ
)
=
P
(
X
,
Z
)
P
(
X
)
=
∏
i
=
1
N
p
z
(
i
)
⋅
N
(
x
(
i
)
∣
μ
z
(
i
)
,
Σ
z
(
i
)
)
∑
k
=
1
K
p
k
⋅
N
(
X
∣
μ
k
,
Σ
k
)
\begin{aligned}P(\mathcal X,\mathcal Z \mid \theta) & = P(\mathcal Z = z_j)\cdot P(\mathcal X \mid \mathcal Z = z_j) \\ & = p_{\mathcal Z} \cdot \mathcal N(\mathcal X \mid \mu_{\mathcal Z},\Sigma_{\mathcal Z}) \\ & = \prod_{i=1}^N p_{z^{(i)}} \cdot \mathcal N(x^{(i)} \mid \mu_{z^{(i)}},\Sigma_{z^{(i)}})\\ P(\mathcal Z \mid \mathcal X,\theta) & = \frac{P(\mathcal X,\mathcal Z)}{P(\mathcal X)} \\ & = \frac{\prod_{i=1}^N p_{z^{(i)}} \cdot \mathcal N(x^{(i)} \mid \mu_{z^{(i)}},\Sigma_{z^{(i)}})}{\sum_{k=1}^{\mathcal K} p_k\cdot \mathcal N(\mathcal X \mid \mu_k,\Sigma_k)} \end{aligned}
P(X,Z∣θ)P(Z∣X,θ)=P(Z=zj)⋅P(X∣Z=zj)=pZ⋅N(X∣μZ,ΣZ)=i=1∏Npz(i)⋅N(x(i)∣μz(i),Σz(i))=P(X)P(X,Z)=∑k=1Kpk⋅N(X∣μk,Σk)∏i=1Npz(i)⋅N(x(i)∣μz(i),Σz(i))
求解过程(E步过程)
-
已知 L ( θ , θ ( t ) ) \mathcal L(\theta,\theta^{(t)}) L(θ,θ(t))函数表示如下：
L ( θ , θ ( t ) ) = ∫ Z log P ( X , Z ∣ θ ) ⋅ P ( Z ∣ X , θ ( t ) ) d Z \mathcal L(\theta,\theta^{(t)}) = \int_{\mathcal Z} \log P(\mathcal X,\mathcal Z \mid \theta) \cdot P(\mathcal Z \mid \mathcal X,\theta^{(t)}) d\mathcal Z L(θ,θ(t))=∫ZlogP(X,Z∣θ)⋅P(Z∣X,θ(t))dZ -
将 P ( X , Z ∣ θ ) , P ( Z ∣ X , θ ) P(\mathcal X,\mathcal Z \mid \theta),P(\mathcal Z \mid \mathcal X,\theta) P(X,Z∣θ),P(Z∣X,θ)代入上式：
由于‘高斯混合模型’隐变量Z \mathcal Z Z是离散型参数，因而将∫ \int ∫符号改为∑ \sum ∑符号，并且各样本之间服从’独立同分布‘。
∑ Z log ∏ i = 1 N P ( x ( i ) , z ( i ) ∣ θ ) ⋅ ∏ i = 1 N P ( z ( i ) ∣ x ( i ) , θ ( t ) ) \sum_{\mathcal Z} \log \prod_{i=1}^N P(x^{(i)},z^{(i)} \mid \theta) \cdot \prod_{i=1}^N P(z^{(i)} \mid x^{(i)},\theta^{(t)}) Z∑logi=1∏NP(x(i),z(i)∣θ)⋅i=1∏NP(z(i)∣x(i),θ(t)) -
将 log ∏ i = 1 N P ( x ( i ) , z ( i ) ∣ θ ) \log \prod_{i=1}^N P(x^{(i)},z^{(i)} \mid \theta) log∏i=1NP(x(i),z(i)∣θ)进行变换，并将 ∑ Z \sum_{\mathcal Z} ∑Z展开：
∑ z ( 1 ) , z ( 2 ) , ⋯ , z ( N ) ∑ i = 1 N log P ( x ( i ) , z ( i ) ∣ θ ) ⋅ ∏ i = 1 N P ( z ( i ) ∣ x ( i ) , θ ( t ) ) \sum_{z^{(1)},z^{(2)},\cdots,z^{(N)}} \sum_{i=1}^N \log P(x^{(i)},z^{(i)} \mid \theta) \cdot \prod_{i=1}^N P(z^{(i)} \mid x^{(i)},\theta^{(t)}) z(1),z(2),⋯,z(N)∑i=1∑NlogP(x(i),z(i)∣θ)⋅i=1∏NP(z(i)∣x(i),θ(t))
关于 ∑ z ( 1 ) , z ( 2 ) , ⋯ , z ( N ) \sum_{z^{(1)},z^{(2)},\cdots,z^{(N)}} ∑z(1),z(2),⋯,z(N)这个形式 需要解释一下，我们之前并没有讨论过 z ( i ) ( i = 1 , 2 , ⋯ , N ) z^{(i)}(i=1,2,\cdots,N) z(i)(i=1,2,⋯,N)到底是什么，只是知道每个样本下 x ( i ) x^{(i)} x(i)均对应一个隐变量 z ( i ) z^{(i)} z(i)。z ( i ) z^{(i)} z(i)不是一个具体数值，而是一个向量。它表示样本 x ( i ) x^{(i)} x(i)“可能属于的高斯分布”所组成的向量。示例：
依然假设样本空间内一共包含 K \mathcal K K个高斯分布，样本 x ( 1 ) x^{(1)} x(1)对应的隐变量 z ( 1 ) z^{(1)} z(1)表示如下：
z ( 1 ) = ( z 1 ( 1 ) , z 2 ( 1 ) , ⋯ , z K ( 1 ) ) T z^{(1)} = (z_1^{(1)},z_2^{(1)},\cdots,z_{\mathcal K}^{(1)})^{T} z(1)=(z1(1),z2(1),⋯,zK(1))T
其中， z k ( 1 ) ( k = 1 , 2 , ⋯ , K ) z_{k}^{(1)}(k=1,2,\cdots,\mathcal K) zk(1)(k=1,2,⋯,K)表示 样本 x ( 1 ) x^{(1)} x(1)可能属于编号为 k k k的高斯分布。注意， z k ( 1 ) z_k^{(1)} zk(1)只表示高斯分布的编号(或者称为离散参数)，它不表示概率。 它是如何表示概率的？结果如下表：z ( 1 ) z^{(1)} z(1) z 1 ( 1 ) z_1^{(1)} z1(1) z 2 ( 1 ) z_2^{(1)} z2(1) ⋯ \cdots ⋯ z K ( 1 ) z_{\mathcal K}^{(1)} zK(1) P ( z ( 1 ) ) P(z^{(1)}) P(z(1)) p 1 ( 1 ) p_1^{(1)} p1(1) p 2 ( 1 ) p_2^{(1)} p2(1) ⋯ \cdots ⋯ p K ( 1 ) p_{\mathcal K}^{(1)} pK(1) p j ( i ) p_j^{(i)} pj(i)是样本点 x ( i ) x^{(i)} x(i)指向编号为 z j z_j zj的隐变量对应的高斯分布 N ( μ j , Σ j ) \mathcal N(\mu_j,\Sigma_j) N(μj,Σj)的概率，而 P ( z ( i ) ) P(z^{(i)}) P(z(i))表示 K \mathcal K K个概率结果组成的向量。用数学语言表达即：
p j ( i ) = P ( x ( i ) → z j ) = P ( x ( i ) ∈ N ( μ j , Σ j ) ) P ( z ( i ) ) = ( p 1 ( i ) , p 2 ( i ) , ⋯ , p K ( i ) ) T p_j^{(i)} = P(x^{(i)} \to z_j) = P(x^{(i)} \in \mathcal N(\mu_j,\Sigma_j)) \\ P(z^{(i)}) = (p_1^{(i)},p_2^{(i)},\cdots,p_{\mathcal K}^{(i)}) ^{T} pj(i)=P(x(i)→zj)=P(x(i)∈N(μj,Σj))P(z(i))=(p1(i),p2(i),⋯,pK(i))T
同样存在这种现象的不仅仅是概率，还有均值、协方差：-
μ
z
(
i
)
\mu_{z^{(i)}}
μz(i)表示样本点
x
(
i
)
x^{(i)}
x(i)对应在
K
\mathcal K
K个高斯分布上的期望结果组成的向量：
μ z ( i ) = ( μ 1 ( i ) ， μ 2 ( i ) , ⋯ , μ K ( i ) ) T \mu_{z^{(i)}} = (\mu_{1}^{(i)}，\mu_2^{(i)}, \cdots, \mu_{\mathcal K}^{(i)})^{T} μz(i)=(μ1(i)，μ2(i),⋯,μK(i))T -
Σ
z
(
i
)
\Sigma_{z^{(i)}}
Σz(i)表示样本点
x
(
i
)
x^{(i)}
x(i)对应在
K
\mathcal K
K个高斯分布上的协方差结果组成的向量：
Σ z ( i ) = ( Σ 1 ( i ) ， Σ 2 ( i ) , ⋯ , Σ K ( i ) ) T \Sigma_{z^{(i)}} = (\Sigma_{1}^{(i)}，\Sigma_2^{(i)}, \cdots, \Sigma_{\mathcal K}^{(i)})^{T} Σz(i)=(Σ1(i)，Σ2(i),⋯,ΣK(i))T
由于 ∑ i = 1 N log P ( x ( i ) , z ( i ) ∣ θ ) \sum_{i=1}^N \log P(x^{(i)},z^{(i)}\mid \theta) ∑i=1NlogP(x(i),z(i)∣θ)中隐变量的形式是 z ( i ) ( i = 1 , 2 , ⋯ , N ) z^{(i)}(i=1,2,\cdots,N) z(i)(i=1,2,⋯,N)而不是 z j ( j = 1 , 2 , ⋯ , K ) z_j(j=1,2,\cdots,\mathcal K) zj(j=1,2,⋯,K)因此对 ∑ Z \sum_{\mathcal Z} ∑Z的展开不是 ∑ z 1 , z 2 , ⋯ , z K \sum_{z_1,z_2,\cdots,z_{\mathcal K}} ∑z1,z2,⋯,zK而是 ∑ z ( 1 ) , z ( 2 ) , ⋯ , z ( N ) \sum_{z^{(1)},z^{(2)},\cdots,z^{(N)}} ∑z(1),z(2),⋯,z(N)。
-
μ
z
(
i
)
\mu_{z^{(i)}}
μz(i)表示样本点
x
(
i
)
x^{(i)}
x(i)对应在
K
\mathcal K
K个高斯分布上的期望结果组成的向量：
-
继续将 ∑ i = 1 N log P ( x ( i ) , z ( i ) ∣ θ ) \sum_{i=1}^N \log P(x^{(i)},z^{(i)}\mid \theta) ∑i=1NlogP(x(i),z(i)∣θ)展开，展开结果如下：
L ( θ , θ ( t ) ) = ∑ z ( 1 ) , z ( 2 ) , ⋯ , z ( N ) [ log P ( x ( 1 ) , z ( 1 ) ∣ θ ) + P ( x ( 2 ) , z ( 2 ) ∣ θ ) + ⋯ + P ( x ( N ) , z ( N ) ∣ θ ) ] ⋅ ∏ i = 1 N P ( z ( i ) ∣ x ( i ) , θ ( t ) ) = [ ∑ z ( 1 ) , z ( 2 ) , ⋯ , z ( N ) log P ( x ( 1 ) , z ( 1 ) ∣ θ ) ∏ i = 1 N P ( z ( i ) ∣ x ( i ) , θ ( t ) ) ] + ⋯ + [ ∑ z ( 1 ) , z ( 2 ) , ⋯ , z ( N ) log P ( x ( N ) , z ( N ) ∣ θ ) ∏ i = 1 N P ( z ( i ) ∣ x ( i ) , θ ( t ) ) ] \begin{aligned} \mathcal L(\theta,\theta^{(t)}) & = \sum_{z^{(1)},z^{(2)},\cdots,z^{(N)}} \left[\log P(x^{(1)},z^{(1)} \mid \theta) + P(x^{(2)},z^{(2)} \mid \theta) + \cdots + P(x^{(N)},z^{(N)} \mid \theta)\right] \cdot \prod_{i=1}^N P(z^{(i)} \mid x^{(i)},\theta^{(t)}) \\ & = \left[\sum_{z^{(1)},z^{(2)},\cdots,z^{(N)}} \log P(x^{(1)},z^{(1)} \mid \theta) \prod_{i=1}^N P(z^{(i)} \mid x^{(i)},\theta^{(t)})\right] + \cdots + \left[\sum_{z^{(1)},z^{(2)},\cdots,z^{(N)}} \log P(x^{(N)},z^{(N)} \mid \theta) \prod_{i=1}^N P(z^{(i)} \mid x^{(i)},\theta^{(t)})\right] \end{aligned} L(θ,θ(t))=z(1),z(2),⋯,z(N)∑[logP(x(1),z(1)∣θ)+P(x(2),z(2)∣θ)+⋯+P(x(N),z(N)∣θ)]⋅i=1∏NP(z(i)∣x(i),θ(t))=⎣ ⎡z(1),z(2),⋯,z(N)∑logP(x(1),z(1)∣θ)i=1∏NP(z(i)∣x(i),θ(t))⎦ ⎤+⋯+⎣ ⎡z(1),z(2),⋯,z(N)∑logP(x(N),z(N)∣θ)i=1∏NP(z(i)∣x(i),θ(t))⎦ ⎤ -
基于上述结果，仅观察第一项：
∑ z ( 1 ) , z ( 2 ) , ⋯ , z ( N ) log P ( x ( 1 ) , z ( 1 ) ∣ θ ) ∏ i = 1 N P ( z ( i ) ∣ x ( i ) , θ ( t ) ) \sum_{z^{(1)},z^{(2)},\cdots,z^{(N)}} \log P(x^{(1)},z^{(1)} \mid \theta) \prod_{i=1}^N P(z^{(i)} \mid x^{(i)},\theta^{(t)}) z(1),z(2),⋯,z(N)∑logP(x(1),z(1)∣θ)i=1∏NP(z(i)∣x(i),θ(t))
观察 ∏ i = 1 N P ( z ( i ) ∣ x ( i ) , θ ( t ) ) \prod_{i=1}^N P(z^{(i)} \mid x^{(i)},\theta^{(t)}) ∏i=1NP(z(i)∣x(i),θ(t))，发现只有第一项 P ( z ( 1 ) ∣ x ( 1 ) , θ ( t ) ) P(z^{(1)} \mid x^{(1)},\theta^{(t)}) P(z(1)∣x(1),θ(t))和 z ( 1 ) z^{(1)} z(1)相关；因此，将上式表示为如下形式：
∑ z ( 1 ) , z ( 2 ) , ⋯ , z ( N ) log P ( x ( 1 ) , z ( 1 ) ∣ θ ) ⋅ P ( z ( 1 ) ∣ x ( 1 ) , θ ( t ) ) ∏ i = 2 N P ( z ( i ) ∣ x ( i ) , θ ( t ) ) = ∑ z ( 1 ) [ log P ( x ( 1 ) , z ( 1 ) ∣ θ ) ⋅ P ( z ( 1 ) ∣ x ( 1 ) , θ ( t ) ) ] ⋅ ∑ z ( 2 ) , ⋯ , z ( N ) [ ∏ i = 2 N P ( z ( i ) ∣ x ( i ) , θ ( t ) ) ] \begin{aligned} & \sum_{z^{(1)},z^{(2)},\cdots,z^{(N)}} \log P(x^{(1)},z^{(1)} \mid \theta) \cdot P(z^{(1)} \mid x^{(1)},\theta^{(t)}) \prod_{i=2}^N P(z^{(i)} \mid x^{(i)},\theta^{(t)}) \\ & = \sum_{z^{(1)}} \left[ \log P(x^{(1)},z^{(1)} \mid \theta) \cdot P(z^{(1)} \mid x^{(1)},\theta^{(t)})\right] \cdot \sum_{z^{(2)},\cdots,z^{(N)}} \left[\prod_{i=2}^N P(z^{(i)} \mid x^{(i)},\theta^{(t)})\right] \end{aligned} z(1),z(2),⋯,z(N)∑logP(x(1),z(1)∣θ)⋅P(z(1)∣x(1),θ(t))i=2∏NP(z(i)∣x(i),θ(t))=z(1)∑[logP(x(1),z(1)∣θ)⋅P(z(1)∣x(1),θ(t))]⋅z(2),⋯,z(N)∑[i=2∏NP(z(i)∣x(i),θ(t))]
观察 ∑ z ( 2 ) , ⋯ , z ( N ) [ ∏ i = 2 N P ( z ( i ) ∣ x ( i ) , θ ( t ) ) ] \sum_{z^{(2)},\cdots,z^{(N)}} \left[\prod_{i=2}^N P(z^{(i)} \mid x^{(i)},\theta^{(t)})\right] ∑z(2),⋯,z(N)[∏i=2NP(z(i)∣x(i),θ(t))]，它可以展开成如下形式：
∑ z ( 2 ) , ⋯ , z ( N ) [ ∏ i = 2 N P ( z ( i ) ∣ x ( i ) , θ ( t ) ) ] = ∑ z ( 2 ) , ⋯ , z ( N ) [ P ( z ( 2 ) ∣ x ( 2 ) , θ ( t ) ) × ⋯ × P ( z ( N ) ∣ x ( N ) , θ ( t ) ) ] = ∑ z ( 2 ) P ( z ( 2 ) ∣ x ( 2 ) , θ ( t ) ) × ⋯ × ∑ z ( N ) P ( z ( N ) ∣ x ( N ) , θ ( t ) ) \begin{aligned} \sum_{z^{(2)},\cdots,z^{(N)}} \left[\prod_{i=2}^N P(z^{(i)} \mid x^{(i)},\theta^{(t)})\right] & = \sum_{z^{(2),\cdots,z^{(N)}}} \left[P(z^{(2)} \mid x^{(2)},\theta^{(t)}) \times \cdots \times P(z^{(N)} \mid x^{(N)},\theta^{(t)})\right] \\ & = \sum_{z^{(2)}} P(z^{(2)} \mid x^{(2)},\theta^{(t)}) \times \cdots \times \sum_{z^{(N)}}P(z^{(N)} \mid x^{(N)},\theta^{(t)}) \end{aligned} z(2),⋯,z(N)∑[i=2∏NP(z(i)∣x(i),θ(t))]=z(2),⋯,z(N)∑[P(z(2)∣x(2),θ(t))×⋯×P(z(N)∣x(N),θ(t))]=z(2)∑P(z(2)∣x(2),θ(t))×⋯×z(N)∑P(z(N)∣x(N),θ(t))
上述结果的任意一项 ∑ z ( j ) P ( z ( j ) ∣ x ( j ) , θ ( t ) ) ( j = 2 , ⋯ , N ) \sum_{z^{(j)}} P(z^{(j)} \mid x^{(j)},\theta^{(t)}) (j=2,\cdots,N) ∑z(j)P(z(j)∣x(j),θ(t))(j=2,⋯,N)都是 基于离散型变量的概率密度积分，因此则有：
∑ z ( 2 ) P ( z ( 2 ) ∣ x ( 2 ) , θ ( t ) ) = ⋯ = ∑ z ( N ) P ( z ( N ) ∣ x ( N ) , θ ( t ) ) = 1 ∑ z ( 2 ) , ⋯ , z ( N ) [ ∏ i = 2 N P ( z ( i ) ∣ x ( i ) , θ ( t ) ) ] = 1 × ⋯ × 1 = 1 \sum_{z^{(2)}} P(z^{(2)} \mid x^{(2)},\theta^{(t)}) = \cdots =\sum_{z^{(N)}} P(z^{(N)} \mid x^{(N)},\theta^{(t)}) = 1 \\ \sum_{z^{(2)},\cdots,z^{(N)}} \left[\prod_{i=2}^N P(z^{(i)} \mid x^{(i)},\theta^{(t)})\right] = 1 \times \cdots \times 1 = 1 z(2)∑P(z(2)∣x(2),θ(t))=⋯=z(N)∑P(z(N)∣x(N),θ(t))=1z(2),⋯,z(N)∑[i=2∏NP(z(i)∣x(i),θ(t))]=1×⋯×1=1
因此，被观察的第一项 结果如下：
∑ z ( 1 ) , z ( 2 ) , ⋯ , z ( N ) log P ( x ( 1 ) , z ( 1 ) ∣ θ ) ∏ i = 1 N P ( z ( i ) ∣ x ( i ) , θ ( t ) ) = ∑ z ( 1 ) [ log P ( x ( 1 ) , z ( 1 ) ∣ θ ) ⋅ P ( z ( 1 ) ∣ x ( 1 ) , θ ( t ) ) ] ⋅ 1 = ∑ z ( 1 ) [ log P ( x ( 1 ) , z ( 1 ) ∣ θ ) ⋅ P ( z ( 1 ) ∣ x ( 1 ) , θ ( t ) ) ] \begin{aligned} & \sum_{z^{(1)},z^{(2)},\cdots,z^{(N)}} \log P(x^{(1)},z^{(1)} \mid \theta) \prod_{i=1}^N P(z^{(i)} \mid x^{(i)},\theta^{(t)}) \\ & = \sum_{z^{(1)}} \left[ \log P(x^{(1)},z^{(1)} \mid \theta) \cdot P(z^{(1)} \mid x^{(1)},\theta^{(t)})\right] \cdot 1 \\ & = \sum_{z^{(1)}} \left[ \log P(x^{(1)},z^{(1)} \mid \theta) \cdot P(z^{(1)} \mid x^{(1)},\theta^{(t)})\right] \end{aligned} z(1),z(2),⋯,z(N)∑logP(x(1),z(1)∣θ)i=1∏NP(z(i)∣x(i),θ(t))=z(1)∑[logP(x(1),z(1)∣θ)⋅P(z(1)∣x(1),θ(t))]⋅1=z(1)∑[logP(x(1),z(1)∣θ)⋅P(z(1)∣x(1),θ(t))] -
基于上一步骤， L ( θ , θ ( t ) ) \mathcal L(\theta,\theta^{(t)}) L(θ,θ(t))可表示为如下形式：
L ( θ , θ ( t ) ) = ∑ z ( 1 ) [ log P ( x ( 1 ) , z ( 1 ) ∣ θ ) ⋅ P ( z ( 1 ) ∣ x ( 1 ) , θ ( t ) ) ] + ⋯ + ∑ z ( N ) [ log P ( x ( N ) , z ( N ) ∣ θ ) ⋅ P ( z ( N ) ∣ x ( N ) , θ ( t ) ) ] = ∑ i = 1 N ∑ z ( i ) [ log P ( x ( i ) , z ( i ) ∣ θ ) ⋅ P ( z ( i ) ∣ x ( i ) , θ ( t ) ) ] \begin{aligned} \mathcal L(\theta,\theta^{(t)}) & = \sum_{z^{(1)}} \left[\log P(x^{(1)},z^{(1)} \mid \theta) \cdot P(z^{(1)} \mid x^{(1)},\theta^{(t)})\right] + \cdots + \sum_{z^{(N)}} \left[\log P(x^{(N)},z^{(N)} \mid \theta) \cdot P(z^{(N)} \mid x^{(N)},\theta^{(t)})\right] \\ & = \sum_{i=1}^N \sum_{z^{(i)}}\left[\log P(x^{(i)},z^{(i)} \mid \theta) \cdot P(z^{(i)} \mid x^{(i)},\theta^{(t)})\right] \end{aligned} L(θ,θ(t))=z(1)∑[logP(x(1),z(1)∣θ)⋅P(z(1)∣x(1),θ(t))]+⋯+z(N)∑[logP(x(N),z(N)∣θ)⋅P(z(N)∣x(N),θ(t))]=i=1∑Nz(i)∑[logP(x(i),z(i)∣θ)⋅P(z(i)∣x(i),θ(t))]
将场景整理中的对应结果代入，有：
关于P ( z ( i ) ) , μ z ( i ) , Σ z ( i ) P(z^{(i)}),\mu_{z^{(i)}},\Sigma_{z^{(i)}} P(z(i)),μz(i),Σz(i)详见上面黄色字解释。
L ( θ , θ ( t ) ) = ∑ i = 1 N ∑ z ( i ) log P ( z ( i ) ) ⋅ N ( x ( i ) ∣ μ z ( i ) , Σ z ( i ) ) ⋅ P ( z ( i ) ) ⋅ N ( x ( i ) ∣ μ z ( i ) , Σ z ( i ) ) ∑ k = 1 K p k ⋅ N ( x ( i ) ∣ μ k , Σ k ) \mathcal L(\theta,\theta^{(t)}) = \sum_{i=1}^N \sum_{z^{(i)}} \log P(z^{(i)}) \cdot \mathcal N(x^{(i)} \mid \mu_{z^{(i)}},\Sigma_{z^{(i)}}) \cdot \frac{P(z^{(i)}) \cdot \mathcal N(x^{(i)} \mid \mu_{z^{(i)}},\Sigma_{z^{(i)}})}{\sum_{k=1}^{\mathcal K} p_k \cdot \mathcal N(x^{(i)} \mid \mu_k,\Sigma_k)} L(θ,θ(t))=i=1∑Nz(i)∑logP(z(i))⋅N(x(i)∣μz(i),Σz(i))⋅∑k=1Kpk⋅N(x(i)∣μk,Σk)P(z(i))⋅N(x(i)∣μz(i),Σz(i))
至此，使用EM算法对高斯混合模型求解过程的E步求解完毕，下一节将介绍M步的求解过程。
p.s.这节视频中符号表示的信息确实很复杂，要多想~
7922
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
