图形深度学习 ML BASICS(I)

来自Siggraph

Machine Learning

机器学习是计算机科学的一个领域,这个领域使用统计学的方法让计算机系统使用数据学习,而不是提前被设计好的。 'ml-define'

ML种类

  • 监督类
    • 分类问题
      • Digit Recognition
      • Spam Detection
      • Face detection
    • 回归问题
      • Human Face/Pose Estimation
      • Model Estimation
    • Data consolidation(数据合并??)
  • 非监督
    • 聚类(Clustering)
      • Group Points According to X 'clu'
      • Image Segmentation using NCuts 'clu'
    • 降维问题(Dimensionality Reduction)
      • 多种学习融合(Manifold Learning)
  • 弱监督/半监督
    • 一些数据采用监督, 一些数据非监督
  • 强化学习
    • Supervision: sparse reward for a sequence of decisions(监督:对一系列决策的稀疏奖励)

PS: Segmentation + Classification in Real Images

'se+clas' 评价标准: Confusion matrix, ROC curve, precision, recall, etc.


Learning a Function

$y = f_w(x)$, 其中$y$是预测(prediction);$f$是预测方法(method);$w$是参数(parameters);$x$是输入(input).

Learning a Linear Separator/Classifier

'linear' 'sh'

$y = f(w_1x_1 + w_2x_2) = \mathcal H(w_1x_1 + w_2x_2)$

其中$\mathcal H$fixed non-linearity 并且$w_1 ,w_2$是通过学习得到的。

Combining Simple Functions/Classifiers

'multi' 'multi'

Regression

1. Least Squares fitting

  • Assumption:Linear Function
\[y = f_\mathbf w(\mathbf x) = f(\mathbf x, \mathbf w) = \mathbf w^T\mathbf w \\ \mathbf w^T\mathbf w = \langle \mathbf w^T,\mathbf w \rangle = \sum_{d=1}^D \mathbf w_d \mathbf x_d \\ \bf x \in \mathbb R^D, \mathbf w \in \mathbb R_D\]
  • Reminder: Linear Classifier 'rs'

1.Sum of Square Errors (MSE without the mean)

$y^i = \mathbf w^T\mathbf x^i + \epsilon^i$

loss function:Sum of Square Errors

$L(\mathbf w) = \sum_{i=1}^N(\epsilon^i)^2$

展开: $L(w_0, w_2) = \sum_{i=1}^N[y_i - (w_0x^i_0 + w_1x^i_1)]^2$

2.Q : what is the best (or least bad) value of w?

$\bf y = \bf X\bf w + \bf \epsilon$

$L(\bf w) = \epsilon^T\epsilon$

$minmize L(\bf w)$

$L(\bf w) = (\bf y - \bf X\bf w)^T(\bf y - \bf X\bf w) = \mid\mid\bf y - \bf X\bf w\mid\mid^2$

$\nabla L = 2\bf X^T(y - Xw) = 0 \Rightarrow w^* = (X^TX)^{-1}X^Ty$
code

  • 超参数
    • 过拟合和未拟合 'hyper'
    • 调参 'tuning'
    • 交叉验证,选择合适的$\lambda$ 'cross'

2. Nonlinear error function and gradient descent

扩展#1:逻辑回归(Logistic Regression)

使用sigmoidal函数$g(\alpha) = \frac{1}{1 + exp(-\alpha)}$
'sq''duibi'

扩展#2:处理多个类别的分类问题

C个类别:one-of-c coding (or one-hot encoding)
矩阵记号: $\mathbf Y = \begin{bmatrix} \mathbf y^1 \ - \ .\ . \ - \ \mathbf y^2\end{bmatrix} = \begin{bmatrix} \mathbf y_1 \mid … \mid \mathbf y_{\mathbf c}\end{bmatrix}$,其中$\mathbf y_{\mathbf c}=\begin{bmatrix} y_1 \…\ y_c^N\end{bmatrix}$, $W = \begin{bmatrix} \bf w_1 \mid … \mid w_c\end{bmatrix}$
损失函数: $L(W) = \sum^C_{c=1}{(\bf y_c - Xw_c)^T(y_c - Xw_c)}$
Least squares fit: $\bf w_c^* = (X^TX)^{-1}X^Ty_c$

Logistic vs Linear Regression (n > 2 classes)

'img'

交叉熵的梯度
  • $L(\mathbf w) =-\sum_{i=1}^Ny^i\log g(\mathbf w^T \mathbf x^i) + (1-y^i)\log (1-g(\mathbf w^T\mathbf x^i))$
  • $\nabla L(\mathbf w^*) == 0$
  • initialize : $\mathbf x_0$
  • Update: $\mathbf x_{i+1} = x_{i} - \alpha \nabla f(\mathbf x_i)$
XOR 问题

3.感知训练(简单神经网络)

  • 'img'
    多层感知机
  • 输入向量
  • 隐藏层
  • 输出