變分推斷 - 簡書

在github看到這個文章寫的不錯息拜，就轉(zhuǎn)載了掘猿，大家一起學(xué)習：
https://github.com/keithyin/mynotes/tree/master/MachineLearning/algorithms

——————————————————————————————————————————
該作者也推薦了下面一個系列的文章，也寫的不錯：

Variational inference
KL(q∥p) minimization-1
KL(p∥q) minimization-2
VAE variation inference變分推理清爽介紹

——————————————————————————————————————————

Variational Auto-Encoder

大家對貝葉斯公式應(yīng)該都很熟悉
$P(Z|X)=\frac{p(X,Z)}{\int_z p(X,Z=z)dz}$

我們稱 $P(Z|X)$ 為posterior distribution。posterior distribution的計算通常是非常困難的，為什么呢袒炉？
假設(shè) $Z$ 是一個高維的隨機變量，如果要求 $P(Z=z|X=x)$ ,我們不可避免的要計算 $\int_z p(X=x,Z=z)dz$ ,由于 $Z$ 是高維隨機變量樊零，這個積分是相當難算的我磁。

variational inference就是用來計算posterior distribution的。

core idea

variational inference的核心思想包含兩步：

假設(shè)分布 $q(z;\lambda)$ (這個分布是我們搞得定的驻襟，搞不定的就沒意義了)
通過改變分布的參數(shù) $\lambda$ , 使 $q(z;\lambda)$ 靠近 $p(z|x)$

總結(jié)稱一句話就是夺艰，為真實的后驗分布引入了一個參數(shù)話的模型。即：用一個簡單的分布 $q(z;\lambda)$ 擬合復(fù)雜的分布 $p(z|x)$ 塑悼。

這種策略將計算 $p(z|x)$ 的問題轉(zhuǎn)化成優(yōu)化問題了
$\lambda^* = \arg\min_{\lambda}~divergence(p(z|x),q(z;\lambda))$
收斂后劲适，就可以用 $q(z;\lambda)$ 來代替 $p(z|x)$ 了

公式推導(dǎo)

$\begin{aligned} \text{log}P(x) &= \text{log}P(x,z)-\text{log}P(z|x) \\ &=\text{log}\frac{P(x,z)}{Q(z;\lambda)}-\text{log}\frac{P(z|x)}{Q(z;\lambda)} \end{aligned}$
等式的兩邊同時對分布 $Q(z)$ 求期望楷掉，得
$\begin{aligned} \mathbb E_{q(z;\lambda)}\text{log}P(x) &= \mathbb E_{q(z;\lambda)}\text{log}P(x,z)-\mathbb E_{q(z;\lambda)}\text{log}P(z|x), \\ \text{log}P(x)&=\mathbb E_{q(z;\lambda)}\text{log}\frac{p(x,z)}{q(z;\lambda)}-\mathbb E_{q(z;\lambda)}\text{log}\frac{p(z|x)}{q(z;\lambda)}, \\ &=KL(q(z;\lambda)||p(z|x))+\mathbb E_{q(z;\lambda)}\text{log}\frac{p(x,z)}{q(z;\lambda)},\\ \text{log}P(x)&=KL(q(z;\lambda)||p(z|x))+\mathbb E_{q(z;\lambda)}\text{log}\frac{p(x,z)}{q(z;\lambda)}, \end{aligned}$
我們的目標是使 $q(z;\lambda)$ 靠近 $p(z|x)$ 厢蒜，就是 $\min_\lambda KL(q(z;\lambda)||p(z|x))$

由于 $KL(q(z;\lambda)||p(z|x))$ 中包含 $p(z|x)$ ，這項非常難求烹植。將 $\lambda$ 看做變量時斑鸦， $\text{log}P(x)$ 為常量，所以草雕， $\min_\lambda KL(q(z;\lambda)||p(z|x))$ 等價于 $\max_\lambda \mathbb E_{q(z;\lambda)}\text{log}\frac{p(x,z)}{q(z;\lambda)}$ 巷屿。

$\mathbb E_{q(z;\lambda)}[\text{log}p(x,z)-\text{log}q(z;\lambda)]$ 稱為Evidence Lower Bound(ELBO)。

現(xiàn)在墩虹，variational inference的目標變成 $\max_\lambda \mathbb E_{q(z;\lambda)}[\text{log}p(x,z)-\text{log}q(z;\lambda)]$

為什么稱之為ELBO呢嘱巾？
$p(x)$ 一般被稱之為evidence，又因為 $KL(q||p)>=0$ , 所以 $p(x)>=E_{q(z;\lambda)}[\text{log}p(x,z)-\text{log}q(z;\lambda)]$ , 這就是為什么被稱為ELBO诫钓。

ELBO

繼續(xù)看一下ELBO
$\begin{aligned} ELBO(\lambda) &= \mathbb E_{q(z;\lambda)}[\text{log}p(x,z)-\text{log}q(z;\lambda)] \\ &= \mathbb E_{q(z;\lambda)}\text{log}p(x,z) -\mathbb E_{q(z;\lambda)}\text{log}q(z;\lambda)\\ &= \mathbb E_{q(z;\lambda)}\text{log}p(x,z) + H(q) \end{aligned}$
The first term represents an energy. The energy encourages $q$ to focus probability mass where the model puts high probability $p(\mathbf{x}, \mathbf{z})$ . The entropy encourages $q$ to spread probability mass to avoid concentrating to one location.

$q(Z)$

$Z$ 包含K個獨立部分(K 維旬昭，當然，第i維也可能是高維向量)菌湃，我們假設(shè)：
$q(Z;\lambda) = \prod_{k=1}^{K}q_k(Z_k;\lambda_k)$
這個被稱為mean field approximation问拘。（關(guān)于mean field approximationhttps://metacademy.org/graphs/concepts/mean_field）

ELBO則變成
$\begin{aligned} ELBO(\lambda) &= \mathbb E_{q(Z;\lambda)}\text{log}p(X,Z) -\mathbb E_{q(z;\lambda)}\text{log}q(Z;\lambda) \\ &= \int q(Z;\lambda)\text{log}p(X,Z)dZ-\int q(Z;\lambda)\text{log}q(Z;\lambda)dZ\\ &=\int [\prod_{k=1}^{K}q_k(Z_k;\lambda_k)] \text{log}p(X,Z)dZ-\int [\prod_{k=1}^{K}q_k(Z_k;\lambda_k)] \text{log}q(Z;\lambda)dZ \end{aligned}$
第一項為 energy，第二項為H(q)

energy

符號的含義：

$Z = \{Z_j,\overline Z_j \}, \overline Z_j=Z\backslash Z_j$
$\lambda=\{\lambda_j, \overline\lambda_j\}, \overline \lambda_j=\lambda\backslash\lambda_j$

先處理第一項：
$\begin{aligned} &\int \Bigr[\prod_{k=1}^{K}q_k(Z_k;\lambda_k)\Bigr] \text{log}p(X,Z)dZ = \\ &\int_{Z_j}q_j(Z_j;\lambda_j)\int_{ \overline Z_j}\Bigr[\prod_{k=1,k\neq j}^K q_k(Z_k;\lambda_k)\Bigr]\text{log}p(X,Z)d \overline Z_jdZ_j = \\ &\int_{Z_j}q_j(Z_j;\lambda_j)\Bigr[E_{q(\overline Z_j;\overline \lambda_j)}\text{log}p(X,Z)\Bigr]dZ_j=\\ &\int_{Z_j}q_j(Z_j;\lambda_j)\{\log \exp\Bigr[E_{q(\overline Z_j;\overline \lambda_j)}\text{log}p(X,Z)\Bigr]\}dZ_j=\\ &\int_{Z_j}q_j(Z_j;\lambda_j)\Bigr[\log q_j^* (Z_j;\lambda_j)+\log C\Bigr]dZ_j \end{aligned}$
其中 $q_j^* (Z_j;\lambda_j)=\frac{1}{C}\exp[E_{q(\overline Z_j;\overline \lambda_j)}\text{log}p(X,Z)]$ , $C$ 保證 $q_j^* (Z_j;\lambda_j)$ 是一個分布惧所。 $C$ 與分布的參數(shù) $\overline \lambda_j$ 有關(guān)骤坐，與變量無關(guān)！下愈！

H(q)

再處理第二項：
$\begin{aligned} &\int \Bigr[\prod_{k=1}^{K}q_k(Z_k;\lambda_k)\Bigr] \text{log}q(Z;\lambda)dZ = \\ &\int \Bigr[\prod_{k=1}^{K}q_k(Z_k;\lambda_k)\Bigr] \sum_{n=1}^K\text{log}q(Z_n;\lambda)dZ = \\ &\sum_j\int \Bigr[\prod_{k=1}^{K}q_k(Z_k;\lambda_k)\Bigr] \text{log}q(Z_j;\lambda_j)dZ=\\ &\sum_j\int \Bigr[\prod_{k=1}^{K}q_k(Z_k;\lambda_k)\Bigr] \text{log}q(Z_j;\lambda_j)dZ=\\ &\sum_j\int_{Z_j} q_j(Z_j;\lambda_j)\text{log}q(Z_j;\lambda_j)dZ_j\int [\prod_{k=1,k\neq j}^{K}q_k(Z_k;\lambda_k)]d\overline Z_j=\\ &\sum_j\int_{Z_j} q_j(Z_j;\lambda_j)\text{log}q(Z_j;\lambda_j)dZ_j \end{aligned}$

再看ELBO

經(jīng)過上面的處理纽绍，ELBO變?yōu)?br> $\begin{aligned} ELBO &= \int_{Z_i}q_i(Z_i;\lambda_j)\text{log}q_i^* (Z_i;\lambda_i)dZ_i-\sum_j\int_{Z_j} q_j(Z_j;\lambda_j)\text{log}q(Z_j;\lambda_j)dZ_j+\log C\\ &=\{\int_{Z_i}q_i(Z_i;\lambda_j)\text{log}q_i^* (Z_i;\lambda_i)dZ_i-\int_{Z_i} q_i(Z_i;\lambda_j)\text{log}q(Z_i;\lambda_i)dZ_i\} +H(q(\overline Z_i;\overline \lambda_i))+\log C\\ & \end{aligned}$
再看上式 $\{\}$ 中的項：
$\int_{Z_i}q_i(Z_i;\lambda_j)\text{log}q_i^* (Z_i;\lambda_i)dZ_i-\int_{Z_i} q_i(Z_i;\lambda_j)\text{log}q(Z_i;\lambda_i)dZ_i = -KL(q_i(Z_i;\lambda_j)||q_i^* (Z_i;\lambda_i))$
所以ELBO又可以寫成：
$ELBO=-KL(q_i(Z_i;\lambda_j)||q_i^* (Z_i;\lambda_i))+H(q(\overline Z_i;\overline \lambda_i))+\log C$
我們要 $maxmize ELBO$ ，如何更新 $q_i(Z_i;\lambda_i)$ 呢势似？
從
$ELBO=-KL(q_i(Z_i;\lambda_i)||q_i^* (Z_i;\lambda_i))+H(q(\overline Z_i;\overline \lambda_i))+\log C$
可以看出顶岸，當 $q_i(Z_i;\lambda_j)=q_i^* (Z_i;\lambda_i)$ 時腔彰， $KL(q_i(Z_i;\lambda_j)||q_i^* (Z_i;\lambda_i))=0$ 。這時辖佣，ELBO取最大值霹抛。
所以參數(shù)更新策略就變成了
$\begin{aligned} &q_1(Z_1;\lambda_1)=q_1^* (Z_1;\lambda_1)\\ &q_2(Z_2;\lambda_2)=q_2^* (Z_2;\lambda_2)\\ &q_3(Z_3;\lambda_3)=q_3^* (Z_3;\lambda_3)\\ &... \end{aligned}$
關(guān)于 $q_i^* (Z_i;\lambda_i)$
$\begin{aligned} q_i(Z_i;\lambda_i)&=q_i^* (Z_i;\lambda_i)\\ q_i (Z_i;\lambda_i)&=\frac{1}{C}\exp[E_{q(\overline Z_i;\overline \lambda_i)}\text{log}p(X,Z)]\\ &=\frac{1}{C}\exp[E_{q(\overline Z_i;\overline \lambda_i)}\text{log}p(X,Z_i,\overline Z_i)]\\ & \end{aligned}$
$q_i$ 是要更新的節(jié)點， $X$ 是觀測的數(shù)據(jù),由于 Markov Blanket(下面介紹)卷谈，更新公式變成：
$\log(q_i(Z_i;\lambda_i))=\int q(mb(Z_i))\log p(Z_i,mb(Z_i),X)d~mb(Z_i)$
由于式子中和 $Z_i$ 無關(guān)的項都被積分積掉了杯拐，所以寫成了 Markov Blanket 這種形式

Markov Blanket

In machine learning, the Markov blanket for a node $A$ in a Bayesian network is the set of nodes $mb(A)$ composed of $A's$ parents, its children, and its children's other parents. In a Markov random field, the Markov blanket of a node is its set of neighboring nodes.
Every set of nodes in the network is conditionally independent of $A$ when conditioned on the set $mb(A)$ , that is, when conditioned on the Markov blanket of the node $A$ . The probability has the Markov property; formally, for distinct nodes $A$ and $B$ :
$Pr(A|mb(A),B)=Pr(A|mb(A))$
The Markov blanket of a node contains all the variables that shield the node from the rest of the network. This means that the Markov blanket of a node is the only knowledge needed to predict the behavior of that node.

參考資料

https://en.wikipedia.org/wiki/Markov_blanket
http://edwardlib.org/tutorials/inference
http://edwardlib.org/tutorials/variational-inference