Skip to content

Latest commit

 

History

History
562 lines (415 loc) · 14.5 KB

cheatsheet.md

File metadata and controls

562 lines (415 loc) · 14.5 KB
Raw

总结

Math

  1. MLE $$ \theta_{MLE}=\mathop{argmax}\limits _{\theta}\log p(X|\theta)\mathop{=}\limits _{iid}\mathop{argmax}\limits _{\theta}\sum\limits {i=1}^{N}\log p(x{i}|\theta) $$

  2. MAP $$ \theta_{MAP}=\mathop{argmax}\limits _{\theta}p(\theta|X)=\mathop{argmax}\limits _{\theta}p(X|\theta)\cdot p(\theta) $$

  3. Gaussian Distribution $$ \begin{align}&p(x|\mu,\Sigma)=\frac{1}{(2\pi)^{p/2}|\Sigma|^{1/2}}e^{-\frac{1}{2}(x-\mu)^{T}\Sigma^{-1}(x-\mu)}\ &\Delta=(x-\mu)^{T}\Sigma^{-1}(x-\mu)=\sum\limits {i=1}^{p}(x-\mu)^{T}u{i}\frac{1}{\lambda_{i}}u_{i}^{T}(x-\mu)=\sum\limits {i=1}^{p}\frac{y{i}^{2}}{\lambda_{i}} \end{align} $$

  4. 已知 $x\sim\mathcal{N}(\mu,\Sigma), y\sim Ax+b$，有： $$ \begin{align}y\sim\mathcal{N}(A\mu+b, A\Sigma A^T) \end{align} $$

  5. 记 $x=(x_1, x_2,\cdots,x_p)^T=(x_{a,m\times 1}, x_{b,n\times1})^T,\mu=(\mu_{a,m\times1}, \mu_{b,n\times1}),\Sigma=\begin{pmatrix}\Sigma_{aa}&\Sigma_{ab}\\Sigma_{ba}&\Sigma_{bb}\end{pmatrix}$，已知 $x\sim\mathcal{N}(\mu,\Sigma)$，则： $$ \begin{align}&x_a\sim\mathcal{N}(\mu_a,\Sigma_{aa})\ &x_b|x_a\sim\mathcal{N}(\mu_{b|a},\Sigma_{b|a})\ &\mu_{b|a}=\Sigma_{ba}\Sigma_{aa}^{-1}(x_a-\mu_a)+\mu_b\ &\Sigma_{b|a}=\Sigma_{bb}-\Sigma_{ba}\Sigma_{aa}^{-1}\Sigma_{ab} \end{align} $$

Linear Regression

Model

  1. Dataset: $$ \mathcal{D}={(x_1, y_1),(x_2, y_2),\cdots,(x_N, y_N)} $$

  2. Notation: $$ X=(x_1,x_2,\cdots,x_N)^T,Y=(y_1,y_2,\cdots,y_N)^T $$

  3. Model: $$ f(w)=w^Tx $$

Loss Function

  1. 最小二乘误差/高斯噪声的MLE $$ L(w)=\sum\limits_{i=1}^N||w^Tx_i-y_i||^2_2 $$

闭式解

$$ \begin{align}\hat{w}=(X^TX)^{-1}X^TY=X^+Y\\ X=U\Sigma V^T\\ X^+=V\Sigma^{-1}U^T \end{align} $$

正则化

$$ \begin{align} L1-Gaussian \ priori&:\mathop{argmin}\limits_wL(w)+\lambda||w||_1,\lambda\gt0\\ L2-Laplasian\ priori-Sparsity&:\mathop{argmin}\limits_wL(w)+\lambda||w||^2_2,\lambda \gt 0 \end{align} $$

Linear Classification

Hard

PCA

  1. Idea: 在线性模型上加入激活函数

  2. Loss Function:

$$ L(w)=\sum\limits_{x_i\in\mathcal{D}_{wrong}}-y_iw^Tx_i $$

  1. Parameters:

$$ w^{t+1}\leftarrow w^{t}+\lambda y_ix_i $$

Fisher

  1. Idea: 投影，类内小，类间大。

  2. Loss Function: $$ \begin{align}&J(w)=\frac{w^TS_bw}{w^TS_ww}\ &S_b=(\overline{x_{c1}}-\overline{x_{c2}})(\overline{x_{c1}}-\overline{x_{c2}})^T\ &S_w=S_1+S_2 \end{align} $$

  3. 闭式解，投影方向: $$ S_w^{-1}(\overline{x_{c1}}-\overline{x_{c2}}) $$

Soft

判别模型

Logistic Regression

  1. Idea，激活函数: $$ \begin{align}p(C_1|x)&=\frac{1}{1+\exp(-a)}\ a&=w^Tx \end{align} $$

  2. Loss Function(交叉熵): $$ \hat{w}=\mathop{argmax}_wJ(w)=\mathop{argmax}w\sum\limits{i=1}^N(y_i\log p_1+(1-y_i)\log p_0) $$

  3. 解法，SGD $$ J'(w)=\sum\limits_{i=1}^N(y_i-p_1)x_i $$

生成模型

GDA

  1. Model

    1. $y\sim Bernoulli(\phi)$
    2. $x|y=1\sim\mathcal{N}(\mu_1,\Sigma)$
    3. $x|y=0\sim\mathcal{N}(\mu_0,\Sigma)$

  2. MAP $$ \begin{align} &\mathop{argmax}{\phi,\mu_0,\mu_1,\Sigma}\log p(X|Y)p(Y)\nonumber\ &=\mathop{argmax}{\phi,\mu_0,\mu_1,\Sigma}\sum\limits_{i=1}^N((1-y_i)\log\mathcal{N}(\mu_0,\Sigma)+y_i\log \mathcal{N}(\mu_1,\Sigma)+y_i\log\phi+(1-y_i)\log(1-\phi)) \end{align} $$

  3. 解 $$ \begin{align}\phi&=\frac{N_1}{N}\ \mu_1&=\frac{\sum\limits_{i=1}^Ny_ix_i}{N_1}\ \mu_0&=\frac{\sum\limits_{i=1}^N(1-y_i)x_i}{N_0}\ \Sigma&=\frac{N_1S_1+N_2S_2}{N} \end{align} $$

Naive Bayesian

  1. Model, 对单个数据点的各个维度作出限制 $$ x_i\perp x_j|y,\forall\ i\ne j $$

    1. $x_i$ 为连续变量：$p(x_i|y)=\mathcal{N}(\mu_i,\sigma_i^2)$
    2. $x_i$ 为离散变量：类别分布（Categorical）：$p(x_i=i|y)=\theta_i,\sum\limits_{i=1}^K\theta_i=1$
    3. $p(y)=\phi^y(1-\phi)^{1-y}$

  2. 解：和GDA相同

Dimension Reduction

中心化： $$ \begin{align}S &=\frac{1}{N}X^T(E_N-\frac{1}{N}\mathbb{I}{N1}\mathbb{I}{1N})(E_N-\frac{1}{N}\mathbb{I}{N1}\mathbb{I}{1N})^TX\nonumber\ &=\frac{1}{N}X^TH^2X=\frac{1}{N}X^THX \end{align} $$

PCA

  1. Idea: 坐标变换，寻找线性无关的新基矢，取信息损失最小的前几个维度

  2. Loss Function: $$ \begin{align}J &=\sum\limits_{j=1}^qu_j^TSu_j\ ,\ s.t.\ u_j^Tu_j=1 \end{align} $$

  3. 解：

    1. 特征分解法 $$ S=U\Lambda U^T $$

    2. SVD for X/S $$ \begin{align}HX=U\Sigma V^T\ S=\frac{1}{N}V\Sigma^T\Sigma V^T \new\ co=HX\cdot V\end{align} $$

    3. SVD for T $$ \begin{align}T=HXX^TH=U\Sigma\Sigma^TU^T\ new\ co=U\Sigma \end{align} $$

p-PCA

  1. Model: $$ \begin{align} z&\sim\mathcal{N}(\mathbb{O}{q1},\mathbb{I}{qq})\ x&=Wz+\mu+\varepsilon\ \varepsilon&\sim\mathcal{N}(0,\sigma^2\mathbb{I}_{pp}) \end{align} $$

  2. Learning: E-M

  3. Inference: $$ p(z|x)=\mathcal{N}(W^T(WW^T+\sigma^2\mathbb{I})^{-1}(x-\mu),\mathbb{I}-W^T(WW^T+\sigma^2\mathbb{I})^{-1}W) $$

SVM

  1. 强对偶关系：凸优化+（松弛）Slater 条件->强对偶。
  2. 参数求解：KKT条件
    1. 可行域
    2. 互补松弛+梯度为0

Hard-margin

  1. Idea: 最大化间隔

  2. Model: $$ \mathop{argmin}_{w,b}\frac{1}{2}w^Tw\ s.t.\ y_i(w^Tx_i+b)\ge1,i=1,2,\cdots,N $$

  3. 对偶问题 $$ \max_{\lambda}-\frac{1}{2}\sum\limits_{i=1}^N\sum\limits_{j=1}^N\lambda_i\lambda_jy_iy_jx_i^Tx_j+\sum\limits_{i=1}^N\lambda_i,\ s.t.\ \lambda_i\ge0 $$

  4. 模型参数 $$ \hat{w}=\sum\limits_{i=1}^N\lambda_iy_ix_i\ \hat{b}=y_k-w^Tx_k=y_k-\sum\limits_{i=1}^N\lambda_iy_ix_i^Tx_k,\exist k,1-y_k(w^Tx_k+b)=0 $$

Soft-margin

  1. Idea:允许少量错误

  2. Model: $$ error=\sum\limits_{i=1}^N\max{0,1-y_i(w^Tx_i+b)}\ \mathop{argmin}{w,b}\frac{1}{2}w^Tw+C\sum\limits{i=1}^N\xi_i\ s.t.\ y_i(w^Tx_i+b)\ge1-\xi_i,\xi_i\ge0,i=1,2,\cdots,N $$

Kernel

对称的正定函数都可以作为正定核。

Exp Family

  1. 表达式 $$ p(x|\eta)=h(x)\exp(\eta^T\phi(x)-A(\eta))=\frac{1}{\exp(A(\eta))}h(x)\exp(\eta^T\phi(x)) $$

  2. 对数配分函数 $$ \begin{align} A'(\eta)=\mathbb{E}{p(x|\eta)}[\phi(x)]\ A''(\eta)=Var{p(x|\eta)}[\phi(x)] \end{align} $$

  3. 指数族分布满足最大熵定理

PGM

Representation

  1. 有向图 $$ p(x_1,x_2,\cdots,x_p)=\prod\limits_{i=1}^pp(x_i|x_{parent(i)}) $$ D-separation $$ p(x_i|x_{-i})=\frac{p(x)}{\int p(x)dx_{i}}=\frac{\prod\limits_{j=1}^pp(x_j|x_{parents(j)})}{\int\prod\limits_{j=1}^pp(x_j|x_{parents(j)})dx_i}=\frac{p(x_i|x_{parents(i)})p(x_{child(i)}|x_i)}{\int p(x_i|x_{parents(i)})p(x_{child(i)}|x_i)dx_i} $$

  2. 无向图 $$ \begin{align}p(x)=\frac{1}{Z}\prod\limits_{i=1}^{K}\phi(x_{ci})\ Z=\sum\limits_{x\in\mathcal{X}}\prod\limits_{i=1}^{K}\phi(x_{ci})\ \phi(x_{ci})=\exp(-E(x_{ci})) \end{align} $$

  3. 有向转无向

    1. 将每个节点的父节点两两相连
    2. 将有向边替换为无向边

Learning

参数学习-EM

  1. 目的：解决具有隐变量的混合模型的参数估计（极大似然估计）

  2. 参数： $$ \theta_{MLE}=\mathop{argmax}\limits_\theta\log p(x|\theta) $$

  3. 迭代求解： $$ \theta^{t+1}=\mathop{argmax}\limits_{\theta}\int_z\log [p(x,z|\theta)]p(z|x,\theta^t)dz=\mathbb{E}_{z|x,\theta^t}[\log p(x,z|\theta)] $$

  4. 原理 $$ \log p(x|\theta^t)\le\log p(x|\theta^{t+1}) $$

  5. 广义EM

    1. E step： $$ \hat{q}^{t+1}(z)=\mathop{argmax}_q\int_zq^t(z)\log\frac{p(x,z|\theta)}{q^t(z)}dz,fixed\ \theta $$

    2. M step： $$ \hat{\theta}=\mathop{argmax}_\theta \int_zq^{t+1}(z)\log\frac{p(x,z|\theta)}{q^{t+1}(z)}dz,fixed\ \hat{q} $$

Inference

  1. 精确推断

    1. VE

    2. BP $$ m_{j\to i}(i)=\sum\limits_j\phi_j(j)\phi_{ij}(ij)\prod\limits_{k\in Neighbour(j)-i}m_{k\to j}(j) $$

    3. MP $$ m_{j\to i}=\max\limits_{j}\phi_j\phi_{ij}\prod\limits_{k\in Neighbour(j)-i}m_{k\to j} $$

  2. 近似推断

    1. 确定性近似，VI

      1. 变分表达式 $$ \hat{q}(Z)=\mathop{argmax}_{q(Z)}L(q) $$

      2. 平均场近似下的 VI-坐标上升 $$ \mathbb{E}{\prod\limits{i\ne j}q_i(Z_i)}[\log p(X,Z)]=\log \hat{p}(X,Z_j)\ q_j(Z_j)=\hat{p}(X,Z_j) $$

      3. SGVI-变成优化问题，重参数法 $$ \mathop{argmax}{q(Z)}L(q)=\mathop{argmax}{\phi}L(\phi)\ \nabla_\phi L(\phi)=\mathbb{E}{q\phi}[(\nabla_\phi\log q_\phi)(\log p_\theta(x^i,z)-\log q_\phi(z))]\ =\mathbb{E}{p(\varepsilon)}[\nabla_z[\log p\theta(x^i,z)-\log q_\phi(z)]\nabla_\phi g_\phi(\varepsilon,x^i)]\ z=g_\phi(\varepsilon,x^i),\varepsilon\sim p(\varepsilon) $$

    2. 随机性近似

      1. 蒙特卡洛方法采样

        1. CDF 采样

        2. 拒绝采样， $q(z)$，使得 $\forall z_i,Mq(z_i)\ge p(z_i)$，拒绝因子：$\alpha=\frac{p(z^i)}{Mq(z^i)}\le1$

        3. 重要性采样 $$ \mathbb{E}{p(z)}[f(z)]=\int p(z)f(z)dz=\int \frac{p(z)}{q(z)}f(z)q(z)dz\simeq\frac{1}{N}\sum\limits{i=1}^Nf(z_i)\frac{p(z_i)}{q(z_i)} $$

        4. 重要性重采样：重要性采样+重采样

      2. MCMC：构建马尔可夫链概率序列，使其收敛到平稳分布 $p(z)$。

        1. 转移矩阵（提议分布） $$ p(z)\cdot Q_{z\to z^}\alpha(z,z^)=p(z^)\cdot Q_{z^\to z}\alpha(z^,z)\ \alpha(z,z^)=\min{1,\frac{p(z^)Q_{z^\to z}}{p(z)Q_{z\to z^*}}} $$

        2. 算法（MH）：

          1. 通过在0，1之间均匀分布取点 $u$
          2. 生成 $z^\sim Q(z^|z^{i-1})$
          3. 计算 $\alpha$ 值
          4. 如果 $\alpha\ge u$，则 $z^i=z^*$，否则 $z^{i}=z^{i-1}$

      3. Gibbs 采样：给定初始值 $z_1^0,z_2^0,\cdots$在 $t+1$ 时刻，采样 $z_i^{t+1}\sim p(z_i|z_{-i})$，从第一个维度一个个采样。

GMM

  1. Model $$ p(x)=\sum\limits_{k=1}^Kp_k\mathcal{N}(x|\mu_k,\Sigma_k) $$

  2. 求解-EM $$ \begin{align}Q(\theta,\theta^t)&=\sum\limits_z[\log\prod\limits_{i=1}^Np(x_i,z_i|\theta)]\prod \limits_{i=1}^Np(z_i|x_i,\theta^t)\nonumber\ &=\sum\limits_z[\sum\limits_{i=1}^N\log p(x_i,z_i|\theta)]\prod \limits_{i=1}^Np(z_i|x_i,\theta^t)\nonumber\ &=\sum\limits_{i=1}^N\sum\limits_{z_i}\log p(x_i,z_i|\theta)p(z_i|x_i,\theta^t)\nonumber\ &=\sum\limits_{i=1}^N\sum\limits_{z_i}\log p_{z_i}\mathcal{N(x_i|\mu_{z_i},\Sigma_{z_i})}\frac{p_{z_i}^t\mathcal{N}(x_i|\mu_{z_i}^t,\Sigma_{z_i}^t)}{\sum\limits_kp_k^t\mathcal{N}(x_i|\mu_k^t,\Sigma_k^t)} \end{align} $$

    $$ p_k^{t+1}=\frac{1}{N}\sum\limits_{i=1}^Np(z_i=k|x_i,\theta^t) $$

序列模型-HMM，LDS，Particle

  1. 假设：

    1. 齐次 Markov 假设（未来只依赖于当前）： $$ p(i_{t+1}|i_t,i_{t-1},\cdots,i_1,o_t,o_{t-1},\cdots,o_1)=p(i_{t+1}|i_t) $$

    2. 观测独立假设： $$ p(o_t|i_t,i_{t-1},\cdots,i_1,o_{t-1},\cdots,o_1)=p(o_t|i_t) $$

  2. 参数 $$ \lambda=(\pi,A,B) $$

离散线性隐变量-HMM

  1. Evaluation：$p(O|\lambda)$，Forward-Backward 算法 $$ p(O|\lambda)=\sum\limits_{i=1}^Np(O,i_T=q_i|\lambda)=\sum\limits_{i=1}^N\alpha_T(i)=\sum\limits_{i=1}^Nb_i(o_1)\pi_i\beta_1(i)\ \alpha_{t+1}(j)=\sum\limits_{i=1}^Nb_{j}(o_t)a_{ij}\alpha_t(i)\ \beta_t(i)=\sum\limits_{j=1}^Nb_j(o_{t+1})a_{ij}\beta_{t+1}(j) $$

  2. Learning：$\lambda=\mathop{argmax}\limits_{\lambda}p(O|\lambda)$，EM 算法（Baum-Welch） $$ \lambda^{t+1}=\mathop{argmax}\lambda\sum\limits_I\log p(O,I|\lambda)p(O,I|\lambda^t)\=\sum\limits_I[\log \pi{i_1}+\sum\limits_{t=2}^T\log a_{i_{t-1},i_t}+\sum\limits_{t=1}^T\log b_{i_t}(o_t)]p(O,I|\lambda^t) $$

  3. Decoding：$I=\mathop{argmax}\limits_{I}p(I|O,\lambda)$，Viterbi 算法-动态规划 $$ \delta_{t}(j)=\max\limits_{i_1,\cdots,i_{t-1}}p(o_1,\cdots,o_t,i_1,\cdots,i_{t-1},i_t=q_i)\\delta_{t+1}(j)=\max\limits_{1\le i\le N}\delta_t(i)a_{ij}b_j(o_{t+1})\\psi_{t+1}(j)=\mathop{argmax}\limits_{1\le i\le N}\delta_t(i)a_{ij} $$

连续线性隐变量-LDS

  1. Model $$ \begin{align} p(z_t|z_{t-1})&\sim\mathcal{N}(A\cdot z_{t-1}+B,Q)\ p(x_t|z_t)&\sim\mathcal{N}(C\cdot z_t+D,R)\ z_1&\sim\mathcal{N}(\mu_1,\Sigma_1) \end{align} $$

  2. 滤波 $$ p(z_t|x_{1:t})=p(x_{1:t},z_t)/p(x_{1:t})\propto p(x_{1:t},z_t)\=p(x_t|z_t)p(z_t|x_{1:t-1})p(x_{1:t-1})\propto p(x_t|z_t)p(z_t|x_{1:t-1}) $$

  3. 递推求解-线性高斯模型

    1. Prediction $$ p(z_t|x_{1:t-1})=\int_{z_{t-1}}p(z_t|z_{t-1})p(z_{t-1}|x_{1:t-1})dz_{t-1}=\int_{z_{t-1}}\mathcal{N}(Az_{t-1}+B,Q)\mathcal{N}(\mu_{t-1},\Sigma_{t-1})dz_{t-1} $$

    2. Update: $$ p(z_t|x_{1:t})\propto p(x_t|z_t)p(z_t|x_{1:t-1} $$

连续非线性隐变量-粒子滤波

通过采样(SIR)解决： $$ \mathbb{E}[f(z)]=\int_zf(z)p(z)dz=\int_zf(z)\frac{p(z)}{q(z)}q(z)dz=\sum\limits_{i=1}^Nf(z_i)\frac{p(z_i)}{q(z_i)} $$

  1. 采样 $$ w_t^i\propto\frac{p(x_t|z_t)p(z_t|z_{t-1})}{q(z_t|z_{1:t-1},x_{1:t})}w_{t-1}^i\ q(z_t|z_{1:t-1},x_{1:t})=p(z_t|z_{t-1}) $$

  2. 重采样

CRF

  1. PDF $$ p(Y=y|X=x)=\frac{1}{Z(x,\theta)}\exp[\theta^TH(y_t,y_{t-1},x)] $$

  2. 边缘概率 $$ p(y_t=i|x)=\sum\limits_{y_{1:t-1}}\sum\limits_{y_{t+1:T}}\frac{1}{Z}\prod\limits_{t'=1}^T\phi_{t'}(y_{t'-1},y_{t'},x)\ p(y_t=i|x)=\frac{1}{Z}\Delta_l\Delta_r\ \Delta_l=\sum\limits_{y_{1:t-1}}\phi_{1}(y_0,y_1,x)\phi_2(y_1,y_2,x)\cdots\phi_{t-1}(y_{t-2},y_{t-1},x)\phi_t(y_{t-1},y_t=i,x)\ \Delta_r=\sum\limits_{y_{t+1:T}}\phi_{t+1}(y_t=i,y_{t+1},x)\phi_{t+2}(y_{t+1},y_{t+2},x)\cdots\phi_T(y_{T-1},y_T,x) $$

    $$ \alpha_t(i)=\Delta_l=\sum\limits_{j\in S}\phi_t(y_{t-1}=j,y_t=i,x)\alpha_{t-1}(j)\ \Delta_r=\beta_t(i)=\sum\limits_{j\in S}\phi_{t+1}(y_t=i,y_{t+1}=j,x)\beta_{t+1}(j) $$

  3. 学习 $$ \nabla_\lambda L=\sum\limits_{i=1}^N\sum\limits_{t=1}^T[f(y_{t-1},y_t,x^i)-\sum\limits_{y_{t-1}}\sum\limits_{y_t}p(y_{t-1},y_t|x^i)f(y_{t-1},y_t,x^i)] $$